Paper deep dive
HiMA-MDD: A Hierarchical Multi-Agent Harness for Interpretable Multimodal Depression Detection in Clinical Interviews
Ao Chen, Xiaojiang Peng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/25/2026, 7:34:22 AM
Summary
The paper introduces HiMA-MDD, a hierarchical multi-agent harness for interpretable depression detection using multimodal clinical interviews. It addresses the 'hierarchical measurement–coordination gap' by aligning assessment hierarchy with three agent layers: Layer 1 for QA-to-item grounding and evidence routing, Layer 2 for multi-factor symptom reasoning via specialized agents, and Layer 3 for global verification and profile reconstruction. Using Qwen2.5-72B-Instruct, the system produces a PHQ-8 profile and screening decision, demonstrating superior performance on the E-DAIC dataset compared to state-of-the-art baselines.
Entities (11)
Relation Signals (8)
HiMA-MDD → outputs → PHQ-8
confidence 100% · reconstructs the verified PHQ-8 profile... The final item scores then deterministically produce the total score and screening decision.
HiMA-MDD → evaluatedon → E-DAIC
confidence 95% · our experiments on E-DAIC demonstrate that HiMA-MDD outperforms the compared state-of-the-art methods.
HiMA-MDD → usesbackbonemodel → Qwen2.5-72B-Instruct
confidence 95% · Using Qwen2.5-72B-Instruct as the harness backbone
HiMA-MDD → consistsof → Layer 2
confidence 90% · Layer 2 assigns symptom groups to operational factor specialists
HiMA-MDD → consistsof → Layer 1
confidence 90% · aligns this assessment hierarchy with three agent layers... Layer 1 identifies candidate QA-to-item relations
HiMA-MDD → consistsof → Layer 3
confidence 90% · Layer 3 audits the complete provisional profile
HiMA-MDD → outperforms → MDAgents
confidence 85% · HiMA-MDD outperforms the compared state-of-the-art methods... MDAgents (Kim et al. 2024)
HiMA-MDD → outperforms → AgentMental
confidence 85% · HiMA-MDD outperforms the compared state-of-the-art methods... AgentMental (Hu et al. 2026)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Depression assessment from multimodal clinical interviews requires integrating dispersed evidence from multiple symptoms into a coherent PHQ-8 profile. This process is hierarchical: relevant evidence is often sparse and context-dependent within local question-answer exchanges, multiple exchanges jointly support symptom-level judgments, and the final assessment depends on the coherence of the complete symptom profile. Existing LLM systems either process interviews holistically or distribute work across generic agent roles; neither design necessarily provides an explicit orchestration mechanism that coordinates evidence access, item-score authority, bounded feedback, and state recording across these levels. To address this gap, we introduce HiMA-MDD, a hierarchical multi-agent harness that aligns this assessment hierarchy with three agent layers. After non-agentic preprocessing constructs context-preserving multimodal QA units, Layer 1 identifies candidate QA-to-item relations and supports bounded item-grounded evidence routing. Layer 2 assigns symptom groups to operational factor specialists, with one specialist responsible for each provisional item score. Layer 3 audits the complete provisional profile, requests at most one round of targeted revision, and reconstructs the verified PHQ-8 profile. This layered design naturally yields a Hierarchical Evidence Trace, preserves all intermediate evidence, judgments, and revisions for auditability. The final item scores then deterministically produce the total score and screening decision. Using Qwen2.5-72B-Instruct as the harness backbone, our experiments on E-DAIC demonstrate that HiMA-MDD outperforms the compared state-of-the-art methods.
Tags
Links
- Source: https://arxiv.org/abs/2608.21868v1
- Canonical: https://arxiv.org/abs/2608.21868v1
Trouble viewing inline? Open PDF directly →
Full Text
63,906 characters extracted from source content.
Expand or collapse full text
HiMA-MDD: A Hierarchical Multi-Agent Harness for Interpretable Multimodal Depression Detection in Clinical Interviews Ao Chen Xiaojiang Peng Abstract Depression assessment from multimodal clinical interviews requires integrating dispersed evidence from multiple symptoms into a coherent PHQ-8 profile. This process is hierarchical: relevant evidence is often sparse and context-dependent within local question–answer exchanges, multiple exchanges jointly support symptom-level judgments, and the final assessment depends on the coherence of the complete symptom profile. Existing LLM systems either process interviews holistically or distribute work across generic agent roles; neither design necessarily provides an explicit orchestration mechanism that coordinates evidence access, item-score authority, bounded feedback, and state recording across these levels. To address this gap, we introduce HiMA-MDD, a hierarchical multi-agent harness that aligns this assessment hierarchy with three agent layers. After non-agentic preprocessing constructs context-preserving multimodal QA units, Layer 1 identifies candidate QA-to-item relations and supports bounded item-grounded evidence routing. Layer 2 assigns symptom groups to operational factor specialists, with one specialist responsible for each provisional item score. Layer 3 audits the complete provisional profile, requests at most one round of targeted revision, and reconstructs the verified PHQ-8 profile. This layered design naturally yields a Hierarchical Evidence Trace, preserves all intermediate evidence, judgments, and revisions for auditability. The final item scores then deterministically produce the total score and screening decision. Using Qwen2.5-72B-Instruct as the harness backbone, our experiments on E-DAIC demonstrate that HiMA-MDD outperforms the compared state-of-the-art methods. Introduction Depression is a common mental disorder characterized by persistent depressed mood or loss of interest and pleasure, often accompanied by changes in sleep, appetite, energy, concentration, and self-worth. It can disrupt relationships, education, employment, and everyday functioning, and severe episodes may be associated with suicide. The World Health Organization estimates that approximately 332 million people worldwide experience depression, including 5.7% of adults, while substantial gaps in access to mental-health care remain (World Health Organization 2025). Reliable and timely assessment is therefore an important clinical and public-health problem, motivating computational methods that can organize information dispersed throughout clinical interviews. Figure 1: Comparison between holistic interview-level assessment and our HiMA-MDD. Computational approaches to depression assessment include holistic prediction over complete interviews as well as methods that incorporate questionnaire structure, retrieve participant-specific evidence, organize long interview content, or produce symptom-aware outputs (Nguyen et al. 2022; Jung et al. 2024; Mandal et al. 2025; Wang et al. 2026; Rosenman, Hendler, and Wolf 2024; Zhang et al. 2025; Chen et al. 2024; Lyu et al. 2026). More recently, multi-agent systems have divided questioning, response-adequacy assessment, scoring, judgment, and updating among specialized functional roles (Kim et al. 2024; Hu et al. 2026; Bi et al. 2025; Greene et al. 2026). Despite this progress, two related limitations remain. Holistic predictors may leave the connection between local interview context and symptom-level decisions implicit, whereas function-oriented multi-agent pipelines do not necessarily specify which evidence each agent may access or how scoring responsibility should align with the PHQ-8 item structure. More broadly, multimodal clinical interviews, symptom reasoning, and PHQ-8 assessment operate at different levels of granularity, but existing systems do not necessarily make the coordination of evidence access, scoring authority, and global revision explicit across these levels. We refer to this disconnect as the hierarchical measurement–coordination gap. As illustrated in Figure 1, the schematic holistic pathway flattens the interview into a single global prediction, leaving evidence attribution implicit and applying PHQ-8 structure mainly at the output. To bridge this gap, we introduce HiMA-MDD, a measurement-aligned hierarchical multi-agent harness for interpretable depression assessment from multimodal clinical interviews, which maintains evidence–item responsibility from local multimodal exchanges to the final PHQ-8 profile. The harness is an explicit orchestration layer that governs what evidence each agent may access, which item scores it is authorized to produce or revise, how feedback is propagated across levels, and how intermediate states are preserved for auditing. Rather than treating multiple agents as an unconstrained workflow, HiMA-MDD aligns the evidence, agent, and measurement hierarchies through three governed agent roles. The QA-to-Item Grounding Agent identifies candidate relations between local interview exchanges and PHQ-8 items and supports the construction of bounded item-grounded evidence bundles. The Multi-Factor Symptom Reasoning Agent coordinates operational symptom-factor specialists, with one specialist responsible for each provisional item score. The Global Symptom Verification Agent audits the complete provisional profile, requests targeted revisions when needed, and reconstructs the verified eight-item PHQ-8 symptom profile. The final item scores are then summed to obtain the PHQ-8 total score and mapped to a screening decision under the prespecified threshold. Together, the recorded evidence links, intermediate judgments, audit findings, and revisions form a Hierarchical Evidence Trace that makes the evidence-to-profile process inspectable. Our contributions are threefold: • A hierarchical formulation of multimodal depression assessment, connecting local interview evidence, symptom-level judgments, and the complete PHQ-8 profile to identify the hierarchical measurement–coordination gap. • A measurement-aligned hierarchical multi-agent harness that governs evidence access, single-owner provisional scoring, bounded feedback, profile reconstruction, and recorded provenance. • A layer-wise evaluation of harnessed reasoning covering overall performance, evidence access, reasoning granularity, global verification and reconstruction, and trace inspectability. Related Work Multimodal Depression Assessment Early multimodal depression-assessment systems commonly predicted a total score or screening label from complete interviews (Al Hanai, Ghassemi, and Glass 2018; Ringeval et al. 2019). Recent systems have incorporated LLM-derived transcript representations, facial-expression features, and multimodal large language models into depression recognition (Sadeghi et al. 2024; Zhang et al. 2026). Other studies have moved toward finer-grained assessment by using questionnaire items or subscores as prediction targets, completing standardized questionnaires from interview text, or retrieving evidence separately for individual questionnaire items (Mandal et al. 2025; Wang et al. 2026; Rosenman, Hendler, and Wolf 2024; Ravenda et al. 2025). Related approaches organize long interviews through question hierarchies or structural graphs and generate PHQ-aware symptom summaries, evidence, or rationales (Zhang et al. 2025; Jung et al. 2024; Chen et al. 2024; Zheng et al. 2025; Lyu et al. 2026). These studies show the value of preserving interview structure and predicting beyond a single total score. However, symptom-related evidence is often treated as an input representation or retrieval result rather than as part of an explicit control interface that also governs downstream scoring responsibility. Multi-Agent Systems for Depression Assessment Multi-agent systems provide another route to structured mental-health assessment. MDAgents adapts the composition of clinical decision teams to task complexity; AgentMental assigns question generation, response-adequacy assessment, scoring, and information updating to different agents; MAGI coordinates specialized roles around a structured psychiatric interview; and AI Psychiatrist Assistant combines assessment, judging, scoring, and review agents (Kim et al. 2024; Hu et al. 2026; Bi et al. 2025; Greene et al. 2026). These systems show how functional specialization can organize complex assessment workflows. HiMA-MDD addresses a different coordination problem: assessment from a completed interview requires explicit control over which evidence each specialist may access, which provisional item scores it may produce, how global feedback may trigger a bounded revision, and which intermediate states are retained. It therefore uses the PHQ-8 measurement process to govern agent responsibilities rather than treating role specialization alone as the organizing principle. Psychometric Structure of the PHQ Psychometric research distinguishes the summed severity score of a questionnaire from the symptom profile represented by its individual items. PHQ scores are commonly obtained by summing item responses, yet people with the same total can have different symptom profiles (Fried and Nesse 2015a; Fried and Nesse 2015b). Research on the internal structure of the PHQ-9 has examined several alternatives. A one-factor model treats all items as indicators of general depression severity. Correlated two-factor models commonly distinguish cognitive/affective and somatic dimensions, whereas bifactor models retain a general factor alongside more specific dimensions (Lamela et al. 2020; Fischer et al. 2022; Chae, Lee, and Lee 2025). Other work has reported a four-group organization comprising Affective, Somatic, Internalizing, and Sensorimotor symptoms (Tseng et al. 2024). These findings do not establish a single universally accepted structure: the supported organization can vary with the population, instrument, and modeling assumptions. HiMA-MDD therefore does not propose or validate a new PHQ-8 factor model. Instead, it adapts the four-group organization as an operational responsibility map; because PHQ-8 omits the suicidality item, the Internalizing specialist is responsible only for low self-worth. The one-group, two-group, four-group, and item-wise configurations are evaluated as alternative reasoning granularities rather than competing psychometric models. Method Task Formulation and Harness Control Interface We introduce HiMA-MDD, a measurement-aligned hierarchical multi-agent harness for interpretable PHQ-8 assessment from completed multimodal clinical interviews. The harness is an explicit orchestration and control layer that implements an execution contract rather than merely naming a sequence of modules: it governs which evidence each role may inspect, which role is authorized to produce each provisional item score, how global feedback may revise that score, and which state transitions must be retained. The agents operate through these interfaces, while the harness constrains their permitted evidence access, score updates, revisions, and outputs. Given a completed interview X, the core hierarchical harness produces eight raw verified PHQ-8 item scores y^ii=18\ y_i\_i=1^8, the corresponding total score S S, a screening decision c c, and a Hierarchical Evidence Trace T: f(X) f(X) →(y^ii=18,S^,c^,T), →(\ y_i\_i=1^8, S, c,T), (1) S S =∑i=18y^i,c^=[S^≥10], = _i=1^8 y_i, c=I[ S≥ 10], where each y^i∈0,1,2,3 y_i∈\0,1,2,3\. Within the core harness, the total score and screening decision follow the fixed PHQ-8 rule. The complete HiMA-MDD assessment harness additionally exposes an optional supervised calibration interface, described below, which maps the frozen raw state to calibrated item scores and then recomputes the total score and screening decision under the same fixed PHQ-8 rule. We formalize the HiMA-MDD Hierarchical Harness as ℋHiMA=(ℐ,ℛ,Ω,,Γ,),H_HiMA=(I,R, ,V, ,T), (2) where ℐI contains the PHQ-8 item rubrics and ordinal scoring semantics; ℛR defines candidate QA-to-item relations, preliminary evidence-polarity metadata, and the bounded evidence-access policy; Ω maps every item to exactly one provisional score owner; V specifies cross-factor audit, targeted revision, and verified-profile reconstruction; Γ deterministically maps the raw verified item vector to its total score and screening decision; and T specifies the provenance and state transitions retained in T. These interfaces enforce four invariants. First, grounding estimates candidate item relevance and preliminary evidence polarity but does not determine symptom severity or produce item scores. Second, each item has exactly one role authorized to write its provisional score, although evidence may be relevant to multiple items. Third, global feedback is targeted, limited to at most one revision round, and recorded together with the resulting response and score change. Fourth, neither an agent nor the verifier independently generates the total score or screening label; both are deterministic consequences of the eight raw verified item scores. Figure 2 instantiates this contract from interview structuring to assessment output. Non-agentic preprocessing constructs context-preserving multimodal QA units and the PHQ-8 measurement contract. Layer 1 builds candidate QA-to-item relations and item-grounded evidence bundles under ℛR; Layer 2 assigns these bundles to single-owner factor specialists under Ω and produces the provisional profile; and Layer 3 applies the audit, targeted revision, and verified-profile reconstruction policy V. The raw verified item vector is passed through Γ , while T retains the grounding, routed evidence, specialist judgments, audit and revision records, and verified profile as the Hierarchical Evidence Trace. The dashed Post-hoc Item-Score Calibration stage represents an optional supervised, measurement-constrained output-control interface in the complete HiMA-MDD assessment harness. It operates only after the core agentic execution is complete, without rerunning the agents or modifying the Hierarchical Evidence Trace, and recalibrates the eight item scores before the total score and screening decision are recomputed. Figure 2: Overview of the HiMA-MDD assessment framework. Data and Measurement Structuring HiMA-MDD first constructs context-preserving Multimodal QA Units using a question-based structure inspired by HiQuE (Jung et al. 2024). Each interviewer question is paired with all consecutive participant response turns before the next question. The resulting unit is uj=(qj,rj,τj,aj),j∈1,…,J,u_j=(q_j,r_j, _j,a_j), j∈\1,…,J\, (3) where qjq_j is the interviewer question, rjr_j is the grouped participant response, τj _j is the turn and timestamp metadata, and aja_j represents any aligned participant-speech acoustic descriptors. Retaining the interviewer question provides local context for short or elliptical responses, while the participant response remains the primary basis for symptom assessment (Burdisso et al. 2024). The PHQ-8 measurement contract supplies eight item rubrics with 00–33 ordinal scoring semantics over the preceding two weeks, the symptom groups that define specialist responsibility, and the fixed rule for computing the total score and screening decision (Kroenke et al. 2009). These rubrics accompany the routed evidence throughout the reasoning layers. Participant-speech acoustic descriptors are aligned to QA units by timestamp. They combine clip-level eGeMAPS functionals extracted with the eGeMAPSv02 configuration (Eyben, Wöllmer, and Schuller 2010; Eyben et al. 2016) and depression and anxiety estimates computed over longer speech contexts by the KintsugiHealth Depression–Anxiety Model (DAM). The DAM model card reports training and evaluation on approximately 863 hours of speech from 35,000 individuals, collected via phone, tablet, or web app and labeled using clinician-administered or self-reported PHQ-9 and GAD-7; it does not list E-DAIC as a training or evaluation source (Kintsugi Health 2026). The descriptors are verbalized and attached to audio-relevant QA evidence as auxiliary cues for the corresponding symptom judgments. Appendix A describes transcript recovery, feature extraction, and descriptor verbalization. Layer 1: QA-to-Item Grounding The QA-to-Item Grounding Agent converts the Multimodal QA Units into an item-level evidence index. For each unit, an LLM relevance mapper identifies candidate PHQ-8 items, estimates item-specific text relevance, assigns preliminary support, contradiction, or uncertainty metadata, records eligible audio cues, and provides a brief mapping rationale. These signals describe candidate relevance and evidence polarity; they do not determine symptom severity or produce item scores. Because a single exchange may inform several symptoms, the resulting Item Relevance Table supports many-to-many QA-to-item relations. An item-aware router then ranks the indexed units separately for each PHQ-8 item, applies lexical fallback when the initial mapping is empty, removes duplicate records, and limits the evidence supplied downstream. It organizes the selected records into supporting, contradictory, and uncertain evidence categories, producing an Item-Grounded Evidence Bundle for each item. Every record retains its QA identifier, question, response, timing, relevance and polarity metadata, and available audio cues. Layer 1 thereby determines which evidence each specialist can inspect; symptom severity assessment and item scoring begin in Layer 2. Layer 2: Multi-Factor Symptom Reasoning The Multi-Factor Symptom Reasoning Agent is instantiated by four parallel specialists. The Affective Specialist assesses Anhedonia and Depressed mood; the Somatic Specialist assesses Sleep disturbance, Fatigue, and Appetite disturbance; the Internalizing Specialist assesses Low self-worth; and the Sensorimotor Specialist assesses Concentration difficulty and Psychomotor disturbance. This organization is informed by prior analyses of PHQ symptom structure (Tseng et al. 2024; Gunzler et al. 2020). Each specialist receives the PHQ-8 rubrics and Item-Grounded Evidence Bundles for its assigned symptoms and returns provisional item scores, confidence estimates, cited supporting and contradictory evidence, evidence sufficiency, indications of missing frequency or temporal information, and concise judgments. Each item has one specialist responsible for its provisional score, while relevant QA evidence may be shared across items. The four specialist reports are combined into the Provisional PHQ-8 Profile supplied to Layer 3. Layer 3: Global Symptom Verification The Global Symptom Verification Agent receives the Provisional PHQ-8 Profile together with the specialist judgments, cited QA identifiers, and grounding and routing metadata, and applies a three-stage verification procedure. First, Cross-Factor Audit checks for unsupported scores, missing temporal or frequency information, repeated or conflicting evidence, somatic–affective confounding, and text–audio inconsistency; each issue is linked to the affected symptom and supporting QA references. Second, when reconsideration is warranted, a Targeted Revision Request specifies the symptom, responsible specialist, relevant QA identifiers, and issue to reconsider. Each implicated specialist receives its previous report, full assigned evidence bundle, and targeted audit instruction. Specialists that are not implicated are not rerun, whereas other items owned by a rerun specialist are requested to remain stable but may be regenerated. HiMA-MDD permits at most one revision round. Third, Verified Profile Reconstruction reconciles the revised and retained judgments with the audit findings, selected QA evidence, compact interview context, and PHQ-8 rubrics. A centralized LLM aggregator records score changes, addresses remaining cross-factor inconsistencies, and produces the Verified PHQ-8 Profile. The fixed PHQ-8 rule converts the raw verified item scores into S S and c c. Throughout the hierarchy, the harness records QA-to-item relations, routed evidence bundles, specialist judgments, provisional scores, audit findings, revision requests and responses, score changes, and the verified profile. These records form the Hierarchical Evidence Trace, allowing each raw verified item score to be inspected alongside its routed evidence and recorded intermediate judgments. Post-hoc Item-Score Calibration Post-hoc Item-Score Calibration implements the optional supervised output-control interface of the complete HiMA-MDD assessment harness. Motivated by systematic scoring biases that may arise when LLM judgments are mapped to ordinal scale items (Zheng et al. 2023; Hada et al. 2024; Jin et al. 2026), it converts the frozen raw HiMA-MDD state into eight calibrated item scores without rerunning the agent hierarchy or modifying the Hierarchical Evidence Trace. Its inputs are standardized numerical features summarizing the provisional and verified scores, score changes, confidence estimates, evidence retrieval, acoustic metadata, and audit and revision records. A separate Bayesian Ridge regressor is learned for each PHQ-8 item. Each regressor is first fitted on the training partition, and its hyperparameters are selected on the development partition. The selected regressor is then refitted on the combined training and development data and used to calibrate the raw test output for that item. Test labels are used only for final evaluation. Each prediction is rounded to the nearest integer and clipped to the PHQ-8 range of 00–33. The eight calibrated item scores are summed to obtain the final total score, and a total of at least 10 yields a positive screening decision. These calibrated item scores, the resulting total, and the screening decision constitute the complete-system outputs reported for E-DAIC, while the raw outputs are retained for evaluating the core agentic harness. Experiments Dataset. We evaluate HiMA-MDD on E-DAIC and DAIC-WOZ, two multimodal clinical-interview corpora with PHQ-8 labels (Gratch et al. 2014; Ringeval et al. 2019). DAIC-WOZ contains 189 Wizard-of-Oz interviews in which the virtual interviewer Ellie is controlled by a human interviewer. E-DAIC extends this corpus to 275 participants and adds interviews conducted by a fully autonomous AI interviewer; its test set consists entirely of autonomous interviews, providing a more difficult interview condition (Gómez-Zaragozá et al. 2026). Following established evaluation settings (Wang et al. 2026; Sadeghi et al. 2024; Hu et al. 2026), we use the held-out E-DAIC test split for the main evaluation and the fixed DAIC-WOZ development split for the additional robustness analysis. Both evaluations use ASR-derived transcripts and participant-speech audio. Each locally evaluated method predicts the eight PHQ-8 item scores on a 00–33 scale; their sum gives the total score, and a total of at least 10 indicates a positive depression screen. Appendix A provides preprocessing and target details. Baselines. We compared HiMA-MDD with three prompt-based baselines (Zero-Shot, 3-Shot, CoT (Wei et al. 2022)) and two recent multi-agent systems: MDAgents (Kim et al. 2024) and AgentMental (Hu et al. 2026). We reran these baselines on the same E-DAIC test set using Qwen2.5-72B-Instruct (Yang et al. 2024)with the same PHQ-8 rubric and output format, predicting eight item scores (summed to total and screening decision). For AgentMental, we replaced its original participant simulator (DeepSeek-R1-Distill-Qwen-32B) with Qwen2.5-72B-Instruct, feeding the completed interview transcript as the response source. Table 1 additionally includes source-reported E-DAIC results from the multimodal MLlm-DR (Zhang et al. 2026) and from the transcript-based Dep-LLM study, which reports Dep-LLM and several general LLMs (Lyu et al. 2026). Appendix B provides implementation and adaptation details. Evaluation. The primary evaluation focuses on PHQ-8 total-score estimation and the thresholded screening decision. The protocol-matched comparison reports Total MAE and RMSE together with screening accuracy, κ, class-wise F1, and Macro-F1. These outcomes are complementary: total-score error measures aggregate severity estimation, whereas screening metrics test the decision induced by the predicted profile. Class-wise F1 reveals whether aggregate screening performance is dominated by one class. Appendix B defines the metrics. Group Method MAE↓ RMSE↓ Acc.↑ Screening κ↑κ F1[C]↑ F1[D]↑ Macro F1↑ Reported Results MLlm-DR 3.57 4.83 — — — — 0.69 GPT-5.5 — — 0.70 — 0.75 0.61 0.68 Gemini-3.1-Pro — — 0.73 — 0.77 0.68 0.73 Claude-Opus-4.6 — — 0.71 — 0.76 0.65 0.71 DeepSeek-V4 — — 0.71 — 0.78 0.60 0.69 Dep-LLM (Gemma3-12B- Instruct) — — 0.75 — 0.80 0.67 0.73 Re-implemented Baselines Zero-Shot 4.54 6.13 0.73 0.35 0.82 0.48 0.65 3-Shot 4.39 5.98 0.73 0.36 0.81 0.52 0.67 CoT 4.27 5.63 0.75 0.42 0.82 0.59 0.70 MDAgents 3.98 5.22 0.75 0.45 0.81 0.63 0.72 AgentMental 4.57 6.22 0.75 0.50 0.77 0.72 0.75 HiMA-MDD HiMA-MDD (raw) 3.96 4.95∗4.95^* 0.80 0.57 0.85 0.72∗0.72^* 0.78 HiMA-MDD w/ Calibration 3.41∗ 4.57∗ 0.84∗ 0.63 0.88∗ 0.74∗ 0.81 Table 1: Comparison of HiMA-MDD with baselines on E-DAIC.Results marked with * are better than Zero-Shot (p<0.05p<0.05) based on metric-specific one-tailed paired tests. Implementation details. We implemented HiMA-MDD as a stateful LangGraph (LangChain, Inc. 2026) workflow. All local systems used Qwen2.5-72B-Instruct (temperature 0), the same PHQ-8 rubric, and shared evaluation code, with inference accelerated via vLLM (Kwon et al. 2023) on NVIDIA RTX 5880 Ada GPUs. For post-hoc calibration output, eight Bayesian Ridge regressors were fitted on the training set, hyperparameter-tuned on the development set, and applied to frozen raw HiMA-MDD test predictions. The screening threshold was fixed at 10 as common. An asterisk indicates an improvement over Zero-Shot at p<0.05p<0.05; Appendix C provides the statistical-testing details. Main Results Method Total MAE↓ Acc.↑ Screening κ↑κ Macro F1↑ Zero-Shot 3.9143 0.7143 0.2081 0.5536 3-Shot 3.7714 0.7143 0.2457 0.5949 CoT 3.4286 0.7429 0.3046 0.6182 MDAgents 3.0571 0.8000 0.4842 0.7281 AgentMental 3.0571 0.7714 0.5122 0.7552 HiMA-MDD (raw) 3.9714 0.8571 0.6765 0.8381 Table 2: Results on the DAIC-WOZ development split. Bold indicates the best performance. Following the established DAIC-WOZ development-cohort protocol, all rows report raw outputs using ASR-derived transcripts. Table 1 reports PHQ-8 total-score estimation and screening results on the held-out E-DAIC test split. HiMA-MDD + Post-hoc Calibration produces the final system output. It achieves a Total MAE of 3.4107 and a Macro-F1 of 0.8130, improving every reported metric over the raw harness output. These are also the best displayed Total MAE and Macro-F1 values among the source-reported and locally evaluated methods. The source-reported rows provide cross-paper context, while the locally rerun rows form the protocol-matched comparison using the same test split, PHQ-8 scoring protocol, and statistical analysis. Among the locally rerun baselines, CoT performs better than Zero-Shot and 3-Shot on Total MAE, Total RMSE, κ, F1[D], and Macro-F1. MDAgents further reduces total-score error, whereas AgentMental obtains the highest F1[D] and Macro-F1 among the baselines. The raw HiMA-MDD output supports mechanism analysis and uncalibrated comparison; among the uncalibrated local methods, it achieves the lowest Total MAE and RMSE and the highest accuracy, κ, F1[C], and Macro-F1, while AgentMental remains slightly higher on F1[D]. Table 2 reports raw-system results on the DAIC-WOZ development split using the E-DAIC prompts and scoring configuration. HiMA-MDD (raw) achieves the highest accuracy, screening κ, and Macro-F1, increasing Macro-F1 from 0.7552 for the best-performing baseline to 0.8381. Its Total MAE is higher, showing that screening performance and absolute total-score error rank the methods differently. Overall, HiMA-MDD maintains robust screening performance on DAIC-WOZ with ASR-derived transcripts. Variant Acc.↑ κ↑κ F1[C]↑ F1[D]↑ Macro-F1↑ HiMA-MDD 0.8036 0.5686 0.8493 0.7179 0.7836 w/o ① 0.7679 0.4902 0.8219 0.6667 0.7443 w/o ② 0.7857 0.5248 0.8378 0.6842 0.7610 w/o ③ 0.7857 0.5340 0.8333 0.7000 0.7667 Table 3: Component ablations of our HiMA-MDD. All rows report raw outputs w/o calibration. ① Cross-factor audit and targeted revision. ② Centralized verified-profile reconstruction. ③ Acoustic descriptor augmentation. Figure 3: Hierarchical Evidence Trace for E-DAIC participant 716, showing evidence retention after domain review for Depressed Mood and multi-evidence aggregation for Feelings of Failure. Effect of Reasoning Granularity and Evidence Reasoning Configuration Total MAE↓ Total RMSE↓ Screening κ↑κ Macro F1↑ Single Agent 4.3036 5.2627 0.4667 0.7333 Two-Factor 4.3571 5.3352 0.5000 0.7499 Four-Factor (Default) 4.0714 5.0533 0.4902 0.7443 Item-Specific 4.3393 5.3802 0.4563 0.7278 Four-Factor w/ Shared Evidence 4.2500 5.0815 0.4340 0.7169 Table 4: Effect of reasoning granularity and evidence access. Configurations share candidate QA-to-item grounding, backbone, rubric, acoustic descriptors, and test cases, and the comparison ends after Layer 2 symptom reasoning. Table 4 examines how reasoning granularity and evidence access affect performance before global verification. The Single Agent, Two-Factor, Four-Factor, and Item-Specific configurations use one, two, four, and eight reasoning agents, respectively. The Single Agent receives a capped evidence bundle retrieved across all eight items; the factor- and item-based configurations receive bounded evidence for their assigned symptoms; and the shared-evidence condition gives the four Factor Specialists the same capped bundle. Two-Factor Specialists divide responsibility between a cognitive–affective group covering Anhedonia, Depressed mood, Low self-worth, Concentration difficulty, and Psychomotor disturbance, and a somatic group covering Sleep disturbance, Fatigue, and Appetite disturbance (Patel et al. 2019; Gunzler et al. 2020). The four-factor responsibility map separates affective, somatic, internalizing, and sensorimotor responsibilities, following prior analyses of PHQ symptom structure (Tseng et al. 2024). Item-Specific Specialists assign one reasoning agent to each PHQ-8 item, whereas the Single Agent condition scores all items together. The results reveal a granularity trade-off: Four-Factor Specialists obtain the lowest Total MAE and RMSE, Two-Factor Specialists achieve the highest κ and Macro-F1, and Item-Specific Specialists score lower than both configurations on score estimation and screening. Thus, four specialists provide the strongest total-score estimation, whereas two specialists favor screening agreement. With the four-factor responsibility map fixed, the item-grounded bounded-access condition performs better than capped shared access on every reported metric. Because the two access conditions differ in both evidence composition and context length, the supported comparison is between the complete bounded-access and capped shared-access policies. Component Ablations of the HiMA-MDD Harness Table 3 evaluates the contributions of cross-factor audit with targeted revision, centralized verified-profile reconstruction, and acoustic descriptor augmentation to the complete raw harness. All three ablations reduce every screening metric. Removing Cross-Factor Audit and Targeted Revision produces the largest decrease, showing that the audit–revision path contributes most strongly among the tested components. Removing centralized reconstruction also lowers performance, indicating that directly combining specialist outputs is less effective than jointly reconciling the verified profile. The decline without acoustic descriptors shows that participant-speech cues provide useful auxiliary information. Case Study Figure 3 shows how the Hierarchical Evidence Trace connects item scores to distributed interview evidence and subsequent review decisions for participant 716. For Depressed Mood, the Affective Specialist assigns a score of 3 using QA716_020 and QA716_028, which describe depression, hopelessness, and hitting rock bottom. Although a related evidence-overlap request triggers re-evaluation of the affective domain, the item is not directly targeted and its evidence and score are retained; the final aggregator records the same score and rationale. For Feelings of Failure, the Internalizing Specialist combines three evidence spans concerning guilt, regret, harmful decisions, and family disconnection to support a score of 2. The trace records the responsible specialist, cited evidence spans, evidence sufficiency, review scope, and final item rationale, allowing both the retained judgment and the multi-evidence judgment to be inspected against their source responses. Conclusion In this paper, we have presented HiMA-MDD, a hierarchical multi-agent harness for PHQ-8 assessment from completed multimodal clinical interviews. HiMA-MDD has organized assessment across evidence grounding, factor-level symptom reasoning, and global symptom verification, producing item scores, total severity, screening decisions, and a recorded Hierarchical Evidence Trace. Results on E-DAIC support the utility of measurement-aligned governance for organizing PHQ-8 assessment while maintaining an inspectable path from interview evidence to the final symptom profile. References Al Hanai, Ghassemi, and Glass (2018) Al Hanai, T.; Ghassemi, M.; and Glass, J. 2018. Detecting Depression with Audio/Text Sequence Modeling of Interviews. In Proceedings of Interspeech 2018, 1716–1720. Bi et al. (2025) Bi, G.; Chen, Z.; Liu, Z.; Wang, H.; Xiao, X.; Xie, Y.; Zhang, W.; Huang, Y.; Chen, Y.; Peng, L.; and Huang, M. 2025. MAGI: Multi-Agent Guided Interview for Psychiatric Assessment. In Findings of the Association for Computational Linguistics: ACL 2025, 24898–24921. Vienna, Austria: Association for Computational Linguistics. Burdisso et al. (2024) Burdisso, S.; Reyes-Ramírez, E.; Villatoro-tello, E.; Sánchez-Vega, F.; Lopez Monroy, A.; and Motlicek, P. 2024. DAIC-WOZ: On the Validity of Using the Therapist’s Prompts in Automatic Depression Detection from Clinical Interviews. In Proceedings of the 6th Clinical Natural Language Processing Workshop, 82–90. Association for Computational Linguistics. Chae, Lee, and Lee (2025) Chae, D.; Lee, J.; and Lee, E.-H. 2025. Internal Structure of the Patient Health Questionnaire-9: A Systematic Review and Meta-analysis. Asian Nursing Research, 19(1): 1–12. Chen et al. (2024) Chen, Z.; Deng, J.; Zhou, J.; Wu, J.; Qian, T.; and Huang, M. 2024. Depression Detection in Clinical Interviews with LLM-Empowered Structural Element Graph. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 8181–8194. Mexico City, Mexico: Association for Computational Linguistics. Dror et al. (2018) Dror, R.; Baumer, G.; Shlomov, S.; and Reichart, R. 2018. The Hitchhiker’s Guide to Testing Statistical Significance in Natural Language Processing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 1383–1392. Association for Computational Linguistics. Eyben et al. (2016) Eyben, F.; Scherer, K. R.; Schuller, B. W.; Sundberg, J.; André, E.; Busso, C.; Devillers, L. Y.; Epps, J.; Laukka, P.; Narayanan, S. S.; and Truong, K. P. 2016. The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing. IEEE Transactions on Affective Computing, 7(2): 190–202. Eyben, Wöllmer, and Schuller (2010) Eyben, F.; Wöllmer, M.; and Schuller, B. 2010. openSMILE: The Munich Versatile and Fast Open-Source Audio Feature Extractor. In Proceedings of the 18th ACM International Conference on Multimedia, 1459–1462. Fischer et al. (2022) Fischer, F.; Levis, B.; Falk, C.; Sun, Y.; Ioannidis, J. P. A.; Cuijpers, P.; Shrier, I.; Benedetti, A.; Thombs, B. D.; and Depression Screening Data (DEPRESSD) PHQ Collaboration. 2022. Comparison of Different Scoring Methods Based on Latent Variable Models of the PHQ-9: An Individual Participant Data Meta-analysis. Psychological Medicine, 52(15): 3472–3483. Fried and Nesse (2015a) Fried, E. I.; and Nesse, R. M. 2015a. Depression Is Not a Consistent Syndrome: An Investigation of Unique Symptom Patterns in the STAR*D Study. Journal of Affective Disorders, 172: 96–102. Fried and Nesse (2015b) Fried, E. I.; and Nesse, R. M. 2015b. Depression Sum-Scores Don’t Add Up: Why Analyzing Specific Depression Symptoms Is Essential. BMC Medicine, 13: 72. Gómez-Zaragozá et al. (2026) Gómez-Zaragozá, L.; Altozano, A.; Llanes-Jurado, J.; Minissi, M. E.; Alcañiz Raya, M.; and Marín-Morales, J. 2026. Detecting Depression through Speech and Text from Casual Talks with Fully Automated Virtual Humans. Artificial Intelligence in Medicine, 171: 103305. Gratch et al. (2014) Gratch, J.; Artstein, R.; Lucas, G.; Stratou, G.; Scherer, S.; Nazarian, A.; Wood, R.; Boberg, J.; DeVault, D.; Marsella, S.; Traum, D.; Rizzo, S.; and Morency, L.-P. 2014. The Distress Analysis Interview Corpus of Human and Computer Interviews. In Proceedings of the Ninth International Conference on Language Resources and Evaluation, 3123–3128. Greene et al. (2026) Greene, A.; Blair, N.; Mahdipour Aghabagher, S.; Kumari, S.; Schlund, M. W.; Fedorov, A.; Calhoun, V. D.; Li, X.; and Silva, R. F. 2026. AI Psychiatrist Assistant: An LLM-based Multi-Agent System for Depression Assessment from Clinical Interviews. In Proceedings of the Fifth Machine Learning for Health Symposium, volume 297 of Proceedings of Machine Learning Research, 525–542. PMLR. Gunzler et al. (2020) Gunzler, D.; Sehgal, A. R.; Kauffman, K.; Davey, C. H.; Dolata, J.; Figueroa, M.; Huml, A.; Pencak, J.; and Sajatovic, M. 2020. Identify Depressive Phenotypes by Applying RDoC Domains to the PHQ-9. Psychiatry Research, 286: 112872. Hada et al. (2024) Hada, R.; Gumma, V.; de Wynter, A.; Diddee, H.; Ahmed, M.; Choudhury, M.; Bali, K.; and Sitaram, S. 2024. Are Large Language Model-Based Evaluators the Solution to Scaling Up Multilingual Evaluation? In Findings of the Association for Computational Linguistics: EACL 2024, 1051–1070. Association for Computational Linguistics. Hu et al. (2026) Hu, J.; Wang, A.; Xie, Q.; Li, Z.; Ma, H.; and Guo, D. 2026. AgentMental: An Interactive Multi-Agent Framework for Explainable and Adaptive Mental Health Assessment. Proceedings of the AAAI Conference on Artificial Intelligence, 40(37): 31050–31058. Jin et al. (2026) Jin, Z.; Hu, J.; Bi, D.; Zhao, K.; and Yu, H. 2026. Evaluating the Efficacy of AI-Based Interactive Assessments Using Large Language Models for Depression Screening: Development and Usability Study. JMIR Formative Research, 10: e78401. Jung et al. (2024) Jung, J.; Kang, C.; Yoon, J.; Kim, S.; and Han, J. 2024. HiQuE: Hierarchical Question Embedding Network for Multimodal Depression Detection. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 1049–1059. Kim et al. (2024) Kim, Y.; Park, C.; Jeong, H.; Chan, Y. S.; Xu, X.; McDuff, D.; Lee, H.; Ghassemi, M.; Breazeal, C.; and Park, H. W. 2024. MDAgents: An Adaptive Collaboration of LLMs for Medical Decision-Making. In Advances in Neural Information Processing Systems, volume 37, 79410–79452. Kintsugi Health (2026) Kintsugi Health. 2026. Depression–Anxiety Model (DAM). Hugging Face model repository. Accessed 2026-07-14. Kroenke et al. (2009) Kroenke, K.; Strine, T. W.; Spitzer, R. L.; Williams, J. B. W.; Berry, J. T.; and Mokdad, A. H. 2009. The PHQ-8 as a Measure of Current Depression in the General Population. Journal of Affective Disorders, 114(1–3): 163–173. Kwon et al. (2023) Kwon, W.; Li, Z.; Zhuang, S.; Sheng, Y.; Zheng, L.; Yu, C. H.; Gonzalez, J. E.; Zhang, H.; and Stoica, I. 2023. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the 29th Symposium on Operating Systems Principles, 611–626. Lamela et al. (2020) Lamela, D.; Soreira, C.; Matos, P.; and Morais, A. 2020. Systematic Review of the Factor Structure and Measurement Invariance of the Patient Health Questionnaire-9 (PHQ-9) and Validation of the Portuguese Version in Community Settings. Journal of Affective Disorders, 276: 220–233. LangChain, Inc. (2026) LangChain, Inc. 2026. LangGraph 1.2.6. GitHub software release. Released 2026-06-18; accessed 2026-07-20. Lyu et al. (2026) Lyu, Y.; Zhao, X.; Tang, B.; and Jiang, R. 2026. Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning. arXiv:2606.10796. Mandal et al. (2025) Mandal, A.; Atzil-Slonim, D.; Solorio, T.; and Gurevych, I. 2025. Enhancing Depression Detection via Question-wise Modality Fusion. In Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2025), 44–61. Albuquerque, New Mexico: Association for Computational Linguistics. Nguyen et al. (2022) Nguyen, T.; Yates, A.; Zirikly, A.; Desmet, B.; and Cohan, A. 2022. Improving the Generalizability of Depression Detection by Leveraging Clinical Questionnaires. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8446–8459. Dublin, Ireland: Association for Computational Linguistics. OpenAI (2024) OpenAI. 2024. Whisper Large-v3-Turbo Model Card. Hugging Face model repository. Accessed 2026-07-16. Patel et al. (2019) Patel, J. S.; Oh, Y.; Rand, K. L.; Wu, W.; Cyders, M. A.; Kroenke, K.; and Stewart, J. C. 2019. Measurement Invariance of the Patient Health Questionnaire-9 (PHQ-9) Depression Screener in U.S. Adults Across Sex, Race/Ethnicity, and Education Level: NHANES 2005–2016. Depression and Anxiety, 36(9): 813–823. Radford et al. (2023) Radford, A.; Kim, J. W.; Xu, T.; Brockman, G.; McLeavey, C.; and Sutskever, I. 2023. Robust Speech Recognition via Large-Scale Weak Supervision. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, 28492–28518. PMLR. Ravenda et al. (2025) Ravenda, F.; Bahrainian, S. A.; Raballo, A.; Mira, A.; and Kando, N. 2025. Are LLMs Effective Psychological Assessors? Leveraging Adaptive RAG for Interpretable Mental Health Screening through Psychometric Practice. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 8975–8991. Vienna, Austria: Association for Computational Linguistics. Reimers and Gurevych (2019) Reimers, N.; and Gurevych, I. 2019. Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, 3982–3992. Association for Computational Linguistics. Ringeval et al. (2019) Ringeval, F.; Schuller, B.; Valstar, M.; Cummins, N.; Cowie, R.; Tavabi, L.; Schmitt, M.; Alisamir, S.; Amiriparian, S.; Messner, E.-M.; Song, S.; Liu, S.; Zhao, Z.; Mallol-Ragolta, A.; Ren, Z.; Soleymani, M.; and Pantic, M. 2019. AVEC 2019 Workshop and Challenge: State-of-Mind, Detecting Depression with AI, and Cross-Cultural Affect Recognition. In Proceedings of the 9th International on Audio/Visual Emotion Challenge and Workshop, 3–12. Rosenman, Hendler, and Wolf (2024) Rosenman, G.; Hendler, T.; and Wolf, L. 2024. LLM Questionnaire Completion for Automatic Psychiatric Assessment. In Findings of the Association for Computational Linguistics: EMNLP 2024, 403–415. Miami, Florida, USA: Association for Computational Linguistics. Sadeghi et al. (2024) Sadeghi, M.; Richer, R.; Egger, B.; Schindler-Gmelch, L.; Rupp, L. H.; Rahimi, F.; Berking, M.; and Eskofier, B. M. 2024. Harnessing Multimodal Approaches for Depression Detection Using Large Language Models and Facial Expressions. npj Mental Health Research, 3: 66. Sentence-Transformers (2021) Sentence-Transformers. 2021. all-mpnet-base-v2 Model Card. Hugging Face model repository. Revision e8c3b32edf5434bc2275fc9bab85f82640a19130; accessed 2026-07-16. Tseng et al. (2024) Tseng, V. W. S.; Tharp, J. A.; Reiter, J. E.; Ferrer, W.; Hong, D. S.; Doraiswamy, P. M.; Nickels, S.; and Project Baseline Health Study Research Group. 2024. Identifying a Stable and Generalizable Factor Structure of Major Depressive Disorder Across Three Large Longitudinal Cohorts. Psychiatry Research, 333: 115702. Wang et al. (2026) Wang, Z.; Li, B.; Tan, W.; Cao, P.; Wang, Y.; Duan, J.; Wang, F.; and Zaiane, O. 2026. Rethinking Depression Prediction from a Fine-Grained Subscore Modeling Perspective via Multi-Task Learning. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 39659–39672. San Diego, California, United States: Association for Computational Linguistics. Wei et al. (2022) Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E. H.; Le, Q. V.; and Zhou, D. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, volume 35, 24824–24837. World Health Organization (2025) World Health Organization. 2025. Depressive Disorder (Depression). WHO fact sheet, accessed 2026-07-13. Wu et al. (2020) Wu, Y.; Levis, B.; Riehm, K. E.; et al. 2020. Equivalency of the Diagnostic Accuracy of the PHQ-8 and PHQ-9: A Systematic Review and Individual Participant Data Meta-analysis. Psychological Medicine, 50(8): 1368–1380. Yang et al. (2024) Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; Lin, H.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Lin, J.; Dang, K.; Lu, K.; Bao, K.; Yang, K.; Yu, L.; Li, M.; Xue, M.; Zhang, P.; Zhu, Q.; Men, R.; Lin, R.; Li, T.; Tang, T.; Xia, T.; Ren, X.; Ren, X.; Fan, Y.; Su, Y.; Zhang, Y.; Wan, Y.; Liu, Y.; Cui, Z.; Zhang, Z.; and Qiu, Z. 2024. Qwen2.5 Technical Report. arXiv:2412.15115. Zhang et al. (2025) Zhang, L.; Gao, Z.; Zhou, D.; and He, Y. 2025. Explainable Depression Detection in Clinical Interviews with Personalized Retrieval-Augmented Generation. In Findings of the Association for Computational Linguistics: ACL 2025, 9927–9944. Vienna, Austria: Association for Computational Linguistics. Zhang et al. (2026) Zhang, W.; Chen, J.; Zhu, E.; Cheng, W.; Li, Y.; Li, Y.; and Wang, Y. J. 2026. MLlm-DR: Towards Explainable Depression Recognition with MultiModal Large Language Models. ACM Transactions on Multimedia Computing, Communications, and Applications, 22(4): 1–23. Zheng et al. (2023) Zheng, L.; Chiang, W.-L.; Sheng, Y.; Zhuang, S.; Wu, Z.; Zhuang, Y.; Lin, Z.; Li, Z.; Li, D.; Xing, E. P.; Zhang, H.; Gonzalez, J. E.; and Stoica, I. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems, volume 36, 46595–46623. Curran Associates, Inc. Zheng et al. (2025) Zheng, W.; Xie, Q.; Wang, Z.; Yu, J.; and Xia, R. 2025. Towards Explainable Multimodal Depression Recognition for Clinical Interviews. arXiv:2501.16106. Technical Appendix Appendix A: Dataset, Preprocessing, and Targets E-DAIC The Extended Distress Analysis Interview Corpus (E-DAIC) was released with the AVEC 2019 Detecting Depression with AI challenge (Ringeval et al. 2019). It extends DAIC-WOZ with interviews conducted by a fully autonomous virtual interviewer. The corpus contains 275 participant sessions and provides audio, video-derived descriptors, automatic transcripts, and depression labels. The official test partition consists entirely of autonomous-agent sessions. HiMA-MDD uses processed transcripts and participant-speech audio. Partition Participants Role in this study Train 163 Post-hoc calibration fitting and final refit Development 56 Post-hoc calibration selection and final refit; three-shot pool Test 56 Final evaluation only Table 5: Official E-DAIC partitions and their use in this study. The test partition is reserved for final reporting. DAIC-WOZ The Distress Analysis Interview Corpus–Wizard of Oz (DAIC-WOZ) contains clinical interviews with 189 participants conducted by the virtual interviewer Ellie under a Wizard-of-Oz protocol, in which a hidden human interviewer controls the interaction (Gratch et al. 2014). The corpus provides interview transcripts, audio recordings, videos, and PHQ-8 labels. Following the development-cohort evaluation setting used in prior work (Hu et al. 2026), we evaluate the locally implemented methods on the fixed 35-participant development split. Our inputs are ASR-derived transcripts reconstructed from the corresponding E-DAIC audio and participant-speech audio rather than the official DAIC-WOZ transcripts. We retain the E-DAIC experimental configuration without DAIC-specific prompt modification or calibration. E-DAIC remains the primary dataset for held-out test evaluation; the DAIC-WOZ cohort is described in the experimental setting as an additional robustness evaluation. Context-Preserving Multimodal QA Construction The input pipeline adapts the hierarchical-question preprocessing of HiQuE (Jung et al. 2024). Each E-DAIC recording is transcribed into time-stamped segments with Whisper turbo (Radford et al. 2023; OpenAI 2024). We then use the 85-question DAIC-WOZ inventory adopted by HiQuE, including each question’s primary/follow-up designation, to recover the interviewer structure. A Sentence-Transformers all-mpnet-base-v2 encoder embeds each candidate interviewer utterance and each inventory question; the normalized dot product selects the nearest inventory entry (Reimers and Gurevych 2019; Sentence-Transformers 2021). Inventory aliases handle common ASR variants. Question-table matching also repairs segments in which an interviewer question lacks a question mark or appears inside a longer ASR span. Matched question text is separated from adjacent participant text, with approximate timestamps assigned by character position within the original segment. Consecutive spans from the same speaker are merged. Files with unresolved turn boundaries are manually inspected, repaired, and resplit from the original waveform. Finally, every interviewer question is paired with all consecutive participant-response turns until the next question. This produces the context-preserving Multimodal QA Units consumed by the Layer 1 QA-to-Item Grounding Agent. Together with the PHQ-8 measurement contract described below, these procedures implement the non-agentic Stage 1 in Figure 2 of the main paper and precede all agent reasoning. Motivated by evidence that interviewer language can create predictive shortcuts (Burdisso et al. 2024), the procedure preserves interviewer questions as local context while treating participant responses as the primary basis for symptom assessment. Participant-Speech Acoustic Descriptor Alignment Only participant-response audio is analyzed. Recordings are resampled to 16 kHz and cut using the repaired transcript timestamps. The first branch applies openSMILE with FeatureSet.eGeMAPSv02 and FeatureLevel.Functionals, yielding 88 eGeMAPS functionals per clip (Eyben, Wöllmer, and Schuller 2010; Eyben et al. 2016). Each Multimodal QA Unit retains a compact subset covering mean pitch and pitch variability, mean loudness and loudness variability, local jitter, local shimmer, harmonic-to-noise ratio (HNR), and the first four MFCC means. The final text supplied to the Factor Specialists verbalizes pitch variability, loudness, jitter, shimmer, and HNR as coarse low/moderate/high descriptions. The bin boundaries are engineering quantization rules, not clinical cutoffs. Descriptor Low/moderate boundary Moderate/high boundary Pitch variability 0.15 0.45 Mean loudness 0.20 0.80 Local jitter 0.015 0.050 Local shimmer (dB) 0.40 1.20 HNR (dB) 5.0 20.0 Table 6: Engineering thresholds used only to verbalize selected eGeMAPSv02 functionals. They do not define depression severity. The second branch uses the official KintsugiHealth/dam Depression–Anxiety Model (DAM) (Kintsugi Health 2026). DAM uses a fine-tuned Whisper-small.en acoustic backbone with task-specific depression and anxiety heads; its model card recommends at least 30 seconds of single-speaker English audio. Our preprocessing uses a 60-second minimum: consecutive participant-response clips are concatenated in interview order until a chunk contains at least 60 seconds of speech. A final shorter remainder is appended to the preceding chunk; if the participant has less than 60 seconds in total, DAM is skipped. HiMA-MDD retains DAM-derived outputs as auxiliary acoustic descriptors. For the depression head, the quantized output is mapped to three bands: 0 for the model’s PHQ-9 0–9 range, 1 for 10–14, and 2 for 15 or above. The resulting depression tag is attached to every participant answer contained in the corresponding chunk because individual answers are often too short for DAM inference. Participant-level summaries record the maximum label, duration-weighted mean label, and positive/severe chunk proportions. The DAM-derived descriptor supplies longer-context acoustic information, while the eGeMAPSv02 description supplies interpretable clip-level context. Both supplement the text and can be grounded in Layer 1 only to depressed mood, fatigue, concentration difficulty, and psychomotor disturbance; they remain auxiliary rather than decisive. PHQ-8 Measurement Contract and Targets The Patient Health Questionnaire-8 (PHQ-8) measures eight symptom categories over the preceding two weeks: anhedonia, depressed mood, sleep disturbance, fatigue, appetite disturbance, low self-worth, concentration difficulty, and psychomotor disturbance (Kroenke et al. 2009). Each item takes an ordinal value in 0,1,2,3\0,1,2,3\, and the item sum ranges from 0 to 24. A systematic review and individual-participant-data meta-analysis reports comparable diagnostic accuracy for the PHQ-8 and PHQ-9 in screening settings (Wu et al. 2020). In Figure 2 of the main paper, the PHQ-8 measurement contract refers to the combination of these item rubrics and ordinal semantics, the operational responsibility map used by the default Factor Specialists, and the fixed aggregation and threshold rules. The four responsibility groups are an implementation choice for evidence routing and provisional scoring, not a newly validated psychometric structure. Every system predicts the complete eight-item profile. The item scores are summed deterministically, and totals of 10 or above are assigned to the depressed range for screening. The same item-scoring and screening rule is applied to gold and predicted profiles. Appendix B: Baselines and Evaluation Protocol Baseline Adaptation The prompting baselines use the same Qwen2.5-72B-Instruct backbone and temperature-zero decoding as HiMA-MDD (Yang et al. 2024). Zero-Shot directly predicts all eight PHQ-8 items, 3-Shot adds three labeled development examples, and Chain-of-Thought requests an explicit reasoning path before item prediction (Wei et al. 2022). Each baseline is run once per participant transcript. MDAgents (Kim et al. 2024) is adapted from its publicly released repository to receive a text-only PHQ-8 JSON task. AgentMental (Hu et al. 2026) is likewise adapted from its publicly released repository, using its PHQ-8 topic and scoring resources, the completed interview transcript as simulated participant history, and the eight topic scores parsed from its final report. Its agents may retain the native follow-up flow, but each simulated answer is generated from that fixed history rather than obtained as a new observation from the original participant. The adaptation therefore evaluates AgentMental in a fixed-evidence setting rather than preserving the information-acquisition advantage of a real online interview. Both systems complete all 56 cases, and their outputs are normalized to the common eight-item schema. These adaptations align the publicly released repository implementations with the E-DAIC PHQ-8 evaluation protocol. Metrics Let N denote the number of participants and M=8M=8 the number of PHQ-8 items. For participant n and item i, yniy_ni and y^ni y_ni are the gold and predicted ordinal scores. With Sn=∑i=1MyniS_n= _i=1^My_ni and S^n=∑i=1My^ni S_n= _i=1^M y_ni, total-score errors are TotalMAE \ MAE =1N∑n=1N|S^n−Sn|, = 1N _n=1^N| S_n-S_n|, TotalRMSE \ RMSE =1N∑n=1N(S^n−Sn)2. = 1N _n=1^N( S_n-S_n)^2. For screening, cn=[Sn≥10]c_n=I[S_n≥ 10] and c^n=[S^n≥10] c_n=I[ S_n≥ 10], where C and D denote the control-range and depressed-range classes. Accuracy is N−1∑n=1N[cn=c^n]N^-1 _n=1^NI[c_n= c_n]. Cohen’s κ is κ=po−pe1−pe,pe=∑k∈C,Dpkp^k,κ= p_o-p_e1-p_e, p_e= _k∈\C,D\p_k p_k, where pop_o is observed agreement and pkp_k and p^k p_k are the gold and predicted proportions of class k. Item κ applies the same unweighted definition to the four ordinal score categories for each item and then averages across the eight items. For k∈C,Dk∈\C,D\, class-wise precision, recall, and F1 are Pk P_k =TPkTPk+FPk, = TP_kTP_k+FP_k, Rk R_k =TPkTPk+FNk, = TP_kTP_k+FN_k, F1[k] 1[k] =2PkRkPk+Rk. = 2P_kR_kP_k+R_k. We report F1[C]F1[C], F1[D]F1[D], and Macro-F1=(F1[C]+F1[D])/2Macro -F1=(F1[C]+F1[D])/2. A precision, recall, or F1 value with a zero denominator is set to zero. The primary evaluation reports Total MAE, Total RMSE, accuracy, screening κ, both class-wise F1 values, and Macro-F1. Table 1 reports MLlm-DR’s published F1 value of 0.69 in the Macro-F1 column. Protocol-matched comparisons and significance tests use the locally rerun systems. Statistical Tests All star markers in the main comparison use the zero-shot LLM baseline as the sole reference. Total-score errors use one-tailed paired tests over shared test participants, accuracy uses the exact McNemar test, and class-wise F1 uses paired approximate randomization (Dror et al. 2018). These metric-specific tests are applied to aligned predictions. The resulting p-values are exploratory, nominal, and unadjusted for multiple comparisons. Each p-value corresponds to its reported metric. Screening κ and Macro-F1 are outside the star-testing mechanism, and no marker denotes a comparison with MDAgents or AgentMental. Implementation and Reproducibility Details • Primary test protocol: both gold and predicted PHQ-8 totals are thresholded at 10. • Transcript ASR: Whisper turbo with time-stamped segments. • Interview question inventory: 85 DAIC-WOZ questions with primary/follow-up types, adapted from HiQuE preprocessing. • Inventory matching model: Sentence-Transformers all-mpnet-base-v2 with normalized embeddings and cosine-equivalent dot product. • Audio sample rate: 16 kHz mono. • Hand-crafted acoustic representation: openSMILE eGeMAPSv02 functionals, with 88 extracted dimensions and a compact verbalized subset. • Deep acoustic representation: KintsugiHealth/dam, checkpoint dam3.1.ckpt; 30-second non-overlapping internal windows and a 60-second minimum concatenated participant-response chunk in our wrapper. • DAM use in HiMA-MDD: DAM-derived acoustic descriptors are retained as auxiliary input. • Audio-eligible PHQ-8 items: depressed mood, fatigue, concentration difficulty, and psychomotor disturbance. • Agent orchestration: LangGraph (LangChain, Inc. 2026). • Runtime: Python 3.11.8, LangGraph 1.2.6, and scikit-learn 1.9.0. • Backbone for HiMA-MDD and local comparisons: Qwen2.5-72B-Instruct. • Decoding temperature: 0. • Number of complete evaluation runs: one per method and configuration. A HiMA-MDD run contains multiple LLM calls across grounding, parallel Factor Specialists, global audit, any requested revision, and centralized reconstruction. • Model serving and acceleration: vLLM (Kwon et al. 2023). • GPU platform: NVIDIA RTX 5880 Ada Generation with 48 GB memory. • Maximum item-grounded QA records per PHQ-8 item: 8. • Maximum deduplicated records per operational factor: 24. • Maximum targeted specialist revision rounds: 1. • Full post-hoc calibration feature dimension: 376. • Post-hoc score-correction model comparison: standardized Ridge, Bayesian Ridge, and Elastic Net models, with hyperparameters selected on the development partition after fitting on the training partition. • Reported post-hoc calibration model: eight independent Bayesian Ridge regressors fitted on the labeled training and development participants. • Final post-hoc calibration fit: train and development partitions, 219 participants in total. • Final post-hoc calibration application: the selected train+development models are applied to raw test predictions; test labels are used only for final evaluation. • Corrected score conversion: nearest-integer rounding followed by clipping to [0,3][0,3]. • Screening threshold: fixed at 10 and never tuned. The evidence-index cache is keyed by the symptom schema, LLM-prompt version, model, temperature, and Multimodal-QA-Unit content hash. Cached candidate QA-to-item relations are shared across the controlled granularity runs so that the comparison changes the Factor-Specialist configuration and evidence access without rerunning Layer 1 grounding. Appendix C: Significance-Test Details System Metric System value Zero-Shot value Test p Star HiMA-MDD (raw) Total MAE 3.9643 4.5357 Paired t-test 0.1347 No HiMA-MDD (raw) Total RMSE 4.9497 6.1296 Paired t-test on squared error 0.0161 Yes HiMA-MDD (raw) Accuracy 0.8036 0.7321 Exact McNemar 0.1719 No HiMA-MDD (raw) F1[C] 0.8493 0.8193 Paired approximate randomization 0.2113 No HiMA-MDD (raw) F1[D] 0.7179 0.4828 Paired approximate randomization 0.0154 Yes HiMA-MDD + Post-hoc Calibration Total MAE 3.4107 4.5357 Paired t-test 0.0122 Yes HiMA-MDD + Post-hoc Calibration Total RMSE 4.5728 6.1296 Paired t-test on squared error 0.0015 Yes HiMA-MDD + Post-hoc Calibration Accuracy 0.8393 0.7321 Exact McNemar 0.0352 Yes HiMA-MDD + Post-hoc Calibration F1[C] 0.8831 0.8193 Paired approximate randomization 0.0369 Yes HiMA-MDD + Post-hoc Calibration F1[D] 0.7429 0.4828 Paired approximate randomization 0.0120 Yes Table 7: Tests underlying the star markers in the main comparison. All tests use Zero-Shot as the reference and report metric-specific one-tailed p-values. The HiMA-MDD + Post-hoc Calibration condition uses the selected supervised Bayesian Ridge ordinal score-correction model. These exploratory values are nominal and unadjusted for multiple comparisons.