Paper deep dive
Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval
Fatema Ferdous Tamanna, K. M. Merajul Arefin, Md. Abdul Masud
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/22/2026, 3:26:45 AM
Summary
The paper introduces a drift-adaptive two-stream architecture for ICU time-series prediction that structurally decouples physiological and treatment representations to prevent 'biological amnesia' during clinical concept drift. Using MIMIC-IV data, the model confines parameter updates to the treatment stream upon detecting drift, preserving stable physiological representations. It incorporates an attribution-driven Temporal Retrieval-Augmented Generation (RAG) module to ground predictions in era-matched PubMed evidence. Results show that selective adaptation improves discrimination and calibration for vasopressor and septic shock prediction compared to static or fully retrained baselines, while maintaining retrieval consistency.
Entities (12)
Relation Signals (9)
Two-Stream Architecture → decouples → Physiology Stream
confidence 95% · structurally decouples physiological from treatment representations
Two-Stream Architecture → decouples → Treatment Stream
confidence 95% · structurally decouples physiological from treatment representations
MIMIC-IV → usedin → Two-Stream Architecture
confidence 95% · Experiments used 84,792 MIMIC-IV stays
Selective Adaptation → improves → Septic Shock
confidence 90% · Selective adaptation improved vasopressor and septic shock discrimination
Selective Adaptation → improves → Vasopressor
confidence 90% · Selective adaptation improved vasopressor and septic shock discrimination
Temporal RAG → uses → PubMed
confidence 90% · grounds each prediction in patient-specific, era-matched PubMed evidence
Full Retraining → causes → Biological Amnesia
confidence 88% · Full retraining risks biological amnesia... in which stable physiological representations are unintentionally distorted
Temporal RAG → encodes → MedCPT
confidence 85% · Queries are prefixed with label anchors... encoded via MedCPT
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Background: Clinical decision support systems degrade silently as treatment protocols evolve, yet standard adaptation methods treat models as monolithic blocks, unable to distinguish stable patient physiology from shifting institutional practice. Methods: We propose an adaptive clinical intelligence architecture for ICU intervention prediction that structurally decouples physiological from treatment representations, confining parameter updates to the treatment stream upon a dual distributional and accuracy trigger. Automated audit logs record which treatment features drove each adaptation event and how their importance shifted. At inference, an attribution-driven Temporal RAG module grounds each prediction in patient-specific, era-matched PubMed evidence anchored to the patient's dominant physiological features. Experiments used 84,792 MIMIC-IV stays (2008-2022) under strict chronological split. Results: Drift localised entirely to the treatment stream, validating the structural prior. Selective adaptation improved vasopressor and septic shock discrimination and calibration over the static source model. A fully retrained baseline yielded marginally higher aggregate discrimination but missed 26 septic shock cases the framework correctly identified, with none in the reverse direction; retrieval consistency with the pre-adaptation source model was preserved by the framework but degraded substantially in the retrained baseline. Conclusions: Structurally constraining adaptation to drifting components while preserving stable physiological representations enables clinical AI to evolve with practice without distorting learned patient biology. This architecture offers a template for governable, interpretable deployment of adaptive models in high-stakes clinical environments.
Tags
Links
- Source: https://arxiv.org/abs/2607.19020v1
- Canonical: https://arxiv.org/abs/2607.19020v1
Trouble viewing inline? Open PDF directly →
Full Text
52,443 characters extracted from source content.
Expand or collapse full text
Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval Fatema Ferdous Tamanna Corresponding author: tamanna16@cse.pstu.ac.bd Dept. of Computer Science and Information Technology, Patuakhali Science and Technology University, Bangladesh K. M. Merajul Arefin Dept. of Computer Science and Engineering, University of Dhaka, Bangladesh Md. Abdul Masud Dept. of Computer Science and Information Technology, Patuakhali Science and Technology University, Bangladesh Abstract Background: Clinical decision support systems degrade silently as treatment protocols evolve, yet standard adaptation methods treat models as monolithic blocks, unable to distinguish stable patient physiology from shifting institutional practice. Methods: We propose an adaptive clinical intelligence architecture for ICU intervention prediction that structurally decouples physiological from treatment representations, confining parameter updates to the treatment stream upon a dual distributional and accuracy trigger. Automated audit logs record which treatment features drove each adaptation event and how their importance shifted. At inference, an attribution-driven Temporal RAG module grounds each prediction in patient-specific, era-matched PubMed evidence anchored to the patient’s dominant physiological features. Experiments used 84,792 MIMIC-IV stays (2008–2022) under strict chronological split. Results: Drift localised entirely to the treatment stream, validating the structural prior. Selective adaptation improved vasopressor and septic shock discrimination and calibration over the static source model. A fully retrained baseline yielded marginally higher aggregate discrimination but missed 26 septic shock cases the framework correctly identified, with none in the reverse direction; retrieval consistency with the pre-adaptation source model was preserved by the framework but degraded substantially in the retrained baseline. Conclusions: Structurally constraining adaptation to drifting components while preserving stable physiological representations enables clinical AI to evolve with practice without distorting learned patient biology. This architecture offers a template for governable, interpretable deployment of adaptive models in high-stakes clinical environments. Keywords: Clinical concept drift, continual learning, clinical decision support systems, explainable AI, retrieval-augmented generation 1 Introduction Clinical decision support systems (CDSS) deployed in intensive care units face a fundamental challenge: the clinical world does not hold still, but deployed models do. Treatment protocols evolve in response to growing evidence and institutional guidelines. The Surviving Sepsis Campaign regularly updates its vasopressor and fluid recommendations [1, 2], sedation practice shifted after adoption of the ABCDEF bundle [3], and the COVID-19 pandemic compressed years of protocol evolution into months [4, 1]. Such drift can silently erode model calibration without triggering an obvious failure signal. This mismatch between training and deployment distributions is concept drift [5, 6]. Its clinical impact is well documented: performance drops in ICU models are driven by shifts in coding and treatments rather than patient biology [7], and single retrospective splits structurally underestimate real-world degradation [8, 9, 10]. Current solutions fall short. Full retraining risks biological amnesia—a domain-specific manifestation of catastrophic forgetting [11] in which stable physiological representations are unintentionally distorted as the model over-indexes on shifting protocols—and creates governance problems when adaptations are unmonitored [12]. Transfer learning [13, 14] and online learning [15] treat the model as a monolithic block, ignoring a fundamental ICU asymmetry: patient physiology is governed by stable human biology, while treatment patterns reflect mutable institutional practice. Furthermore, deployed CDSS often suffer from explanatory staleness, citing outdated guidelines to justify current predictions, and standard prediction tasks frequently suffer from label leakage that obscures true clinical utility. We address this with a drift-adaptive continual learning framework whose contributions are: 1. Ongoing-need label formulation: replaces initiation-only treatment labels with a formulation asking whether treatment is needed during the prediction horizon, eliminating anti-correlation leakage and clinical relevance loss in standard MIMIC-IV labelling. 2. Two-stream architecture with composite drift monitoring and selective adaptation: structurally decouples physiological dynamics (LSTM) from treatment context (MLP), using a composite detector combining PSI, Kolmogorov–Smirnov, and AUROC signals to trigger adaptation confined exclusively to the treatment stream, leaving physiological representations bitwise identical to the source model. 3. Δ -Attribution metric and governance audit logs: a model-class-agnostic measure of feature-importance recalibration that identifies which treatment features drove each adaptation event and operationalises biological amnesia for cross-architecture comparison. 4. Attribution-driven Temporal RAG: derives patient-specific PubMed queries from per-instance Integrated Gradients attributions and conditions retrieval on the detected drift era, structurally guaranteeing evidence consistency via the frozen physiology stream. We validate all four contributions on 84,792 MIMIC-IV stays [18] under a strict chronological split (2008–2022), including formal ablation studies and a clinician-rated retrieval scaffold. To our knowledge, no prior study combines these elements in a single clinically governable pipeline. 2 Methods 2.1 Dataset, Cohort, and Label Formulation All experiments used MIMIC-IV (v3.1) [18], a freely accessible critical care database comprising de-identified electronic health records from Beth Israel Deaconess Medical Center. We extracted adult ICU stays ≥ 14 hours, excluding stays with zero heart-rate variance (recording artefact), yielding 84,792 stays. A strict chronological split was applied: source training (2008–2013; ntrain=35,765n_train=35,765, nval=8,784n_val=8,784) and a chronological test stream (2014–2022; n=40,243n=40,243). This design directly mirrors the conditions under which a deployed CDSS would encounter distributional shift, and differs from the single retrospective splits common in prior MIMIC-based work [7, 10]. During post-drift adaptation a subject-level split (30/10/60%) was used. Cohort characteristics are summarised in Table 1. Standard MIMIC-IV treatment-prediction labels define positive only when treatment initiates in the horizon. This creates two simultaneous pathologies: (i) anti-correlation leakage—patients already on vasopressors get a positive feature but a negative label; (i) clinical relevance loss—continuation and weaning decisions are discarded. We resolve both with the ongoing-need formulation: given observation ∈ℝ6×86X\!∈\!R^6× 86 spanning h0h_0–h5h_5 and a two-hour gap [h6,h7][h_6,h_7] providing sufficient lead time for clinical action [19], yt=[∃h∈[h8,h14]:treatmentt(h)=1]y_t=1\! [∃\,h∈[h_8,h_14]:treatment_t(h)=1 ] (1) generating three binary labels per stay: yvasoy_vaso, yintuby_intub, yshocky_shock. Septic shock is operationalised via all three concurrent Sepsis-3 criteria [20] programmatically identified from MIMIC-IV tables, strictly requiring refractory hypotension: lactate >2mmol/L>2\,mmol/L, active vasopressor, and invasive MAP <65mmHg<65\,mmHg—the most severe subset of Sepsis-3, stricter than the full definition (which includes patients achieving MAP ≥65mmHg≥ 65\,mmHg with vasopressor support). A programmatic audit confirmed absolute temporal isolation across partitions: the intersection of unique subject_ids across every partition combination is exactly zero. Post-drift adaptation used a subject-level split of the 9,602 post-drift stays (30% train / 10% validation / 60% evaluation), yielding 2,885 adaptation training stays, 968 validation stays, and 5,749 held-out evaluation stays (5,749 rather than 5,761 owing to subject-level deduplication removing 12 stays across partition boundaries), plus a 500-stay pre-drift replay buffer. Table 1: Baseline Patient Characteristics Across Temporal Cohorts Characteristic Train/Val (2008–13) Pre-Drift (2014–19) Post-Drift (2020–22) Total Stays (N) 44,549 30,641 9,602 Age, mean± 65.4±16.765.4± 16.7 64.1±16.964.1± 16.9 63.8±16.563.8± 16.5 Male, N (%) 24,291 (54.5) 17,210 (56.2) 5,591 (58.2) White, N (%) 30,909 (69.4) 19,768 (64.5) 5,522 (57.5) Black, N (%) 6,207 (13.9) 2,598 (8.5) 662 (6.9) LOS days, mean± 2.8±4.72.8± 4.7 3.2±5.33.2± 5.3 4.0±7.04.0± 7.0 In-hosp. mortality (%) 9.3 10.0 11.7 Vasopressor, N (%) 7,614 (17.1) 6,067 (19.8) 921 (9.6) Intubation, N (%) 12,098 (27.2) 8,750 (28.6) 2,787 (29.0) Septic Shock, N (%) 830 (1.9) 707 (2.3) 106 (1.1) 2.2 Feature Engineering Features are split into two disjoint sets mirroring the architectural decomposition. The physiological matrix ∈ℝ6×86X\!∈\!R^6× 86 contains hourly vital signs (HR, BP, SpO2, R, temperature), laboratory values (lactate, creatinine, bilirubin, white cell count, platelet count, BUN, glucose, electrolytes, and arterial blood gas parameters), binary observation masks, and derived rolling statistics (baseline-normalised deltas, ratios, and short-horizon mean/SD over 3–6 hour windows). Missing values are handled by within-patient forward-filling followed by median imputation computed exclusively from the training cohort [18]. The treatment vector ∈ℝ12Z\!∈\!R^12 captures static treatment context (crystalloid volume, antibiotic timing, steroid orders, sex, age, medication count) with all direct label proxies excluded to prevent anti-correlation leakage. All continuous treatment features are z-score normalised using training-set statistics. Timing features occurring after the observation window are clipped to a value beyond the observation window length to prevent temporal leakage. 2.3 Two-Stream Neural Architecture The model has three components. The physiology stream processes X through a two-layer LSTM [21] (hidden dim 64, LayerNorm [22]), producing phys∈ℝ64h_phys ^64—frozen after source training in Runs A and B. Layer normalisation is preferred over batch normalisation given the variable-length and irregular-sampling properties of clinical time series. The treatment stream processes z through a two-layer MLP: treat=LayerNorm(ReLU(W2D(ReLU(W1z+b1))+b2))t_treat=LayerNorm(ReLU(W_2D(ReLU(W_1z+b_1))+b_2)) (2) with W1∈ℝ64×12W_1\!∈\!R^64× 12, W2∈ℝ32×64W_2\!∈\!R^32× 64, dropout p=0.3p=0.3. The fusion head concatenates both representations and projects to three outputs: fjoint=[phys;treat]∈ℝ96f_joint=[h_phys;\,t_treat] ^96 (3) through two dense layers (dims 64, 32; ReLU; dropout 0.3/0.225) followed by a linear projection: y^=W5ReLU(W4D2(ReLU(W3D1(fjoint)+b3))+b4)+b5 y=W_5\,ReLU(W_4D_2(ReLU(W_3D_1(f_joint)+b_3))+b_4)+b_5 (4) where D1D_1 and D2D_2 denote the dropout operators for the first and second fusion layers. No activation is used on the final layer, as BCEWithLogitsLoss applies sigmoid internally during training; at inference, probabilities are obtained via σ(y^)σ( y). Training minimises a focal binary cross-entropy loss [23] with per-target positive-class reweighting and label smoothing. For N patients and T=3T=3 targets: ℒ=1NT∑i,t(1−pi,t∗)γ[−αty~i,tlogpi,t−(1−y~i,t)log(1−pi,t)]L= 1NT _i,t(1-p_i,t^*)^γ\! [- _t y_i,t p_i,t-(1- y_i,t) (1-p_i,t) ] (5) where pi,t∗p_i,t^* is the predicted probability of the true class (focal modulating factor), γ=1.5γ=1.5 continuously during both source training and adaptation, y~i,t=yi,t(1−ϵ)+ϵ/2 y_i,t=y_i,t(1-ε)+ε/2 is the label-smoothed target (ϵ=0.02ε=0.02 source for all phases), and αt _t is the target-specific positive-class weight clipped at 20.0. Four neural configurations span the design space: Run A (static source, all frozen); Run B (selective adaptation—proposed: physiology frozen, treatment MLP and fusion head updated); Run C (full adaptation, all layers updated); Run D (single-stream monolithic LSTM baseline: all features concatenated at every timestep; during adaptation only the final output head is updated, mirroring last-layer fine-tuning without structural decomposition). XGBoost [24] trained on flattened physiological statistics concatenated with treatment features (282-dim) serves as the tabular baseline, in both static-source and fully-retrained configurations. 2.4 Dual-Signal Drift Detection and Selective Adaptation Drift is monitored via a dual OR-gate trigger over ten treatment features (excluding age and medication count as stable demographic and administrative covariates). For continuous features, the Population Stability Index (PSI) is computed over 10 quantile bins with Laplace smoothing: PSI=∑i=110(Pi−Qi)ln(Pi/Qi)PSI= _i=1^10(P_i-Q_i) (P_i/Q_i) (6) For binary features, a dedicated binary PSI is computed over outcome categories 0,1\0,1\: PSIbinary=∑k∈0,1(p^kref−p^kcur)ln(p^krefp^kcur)PSI_binary= _k∈\0,1\\! ( p^\,ref_k- p^\,cur_k ) \! ( p^\,ref_k p^\,cur_k ) (7) with proportions clipped to [10−6,1−10−6][10^-6,1-10^-6] for numerical stability. A binary feature is flagged by absolute rate shift >0.05>0.05 or relative shift >20%>20\%. For each feature, three pairwise comparisons are computed: training versus pre-drift (sanity check), training versus post-drift (primary drift signal), and pre-drift versus post-drift (temporal shift within the test stream). Statistical significance is assessed via two-sample Kolmogorov–Smirnov test for continuous features and chi-squared test for binary features (both at α=0.01α=0.01); Cohen’s d effect size is reported for continuous features. Each feature is assigned a categorical drift severity label: stable (PSI<0.10PSI<0.10), minor (0.10≤PSI<0.200.10 <0.20), moderate (0.20≤PSI<0.350.20 <0.35), or severe (PSI≥0.35PSI≥ 0.35). The feature-drift leg combines continuous-feature PSI and binary rate-shifts into a composite distributional score, while factoring in KS-test significance to satisfy a minimum drifted-feature count; the accuracy leg monitors AUROC degradation relative to the source-era baseline; either leg crossing its respective thresholds activates the OR-gate. Drift is triggered if: (1) composite distributional score, computed as PSI¯cont+0.5Δ¯bin PSI_cont+0.5\, _bin, exceeds twice the training-era baseline (floor 0.20); or (2) AUROC drops >0.02>0.02 on any target. Upon trigger, adaptation enforces: ∇θphysℒ=,θm(t+1)←θm(t)−η∇θmℒ,m∈treat,fusion _ _physL=0, _m^(t+1)← _m^(t)-η _ _mL, m∈\treat,fusion\ (8) using a 500-stay pre-drift replay buffer to maintain backward compatibility [11]. Adaptation is run over 5 seeds; median-validation-AUROC model is selected. Automated audit logs record per-feature attribution shifts (via permutation-based feature ablation over a 512-sample adaptation subset) at each adaptation event [25]. 2.5 Δ -Attribution and Biological Amnesia Detection To quantify internal recalibration, we define the Δ -Attribution metric. Global feature importance is: Φj(f,X)=x∼X[|ϕj(f,x)|] _j(f,X)=E_x X[| _j(f,x)|] (9) where ϕj _j is the local attribution score (SHAP for tree ensembles; Integrated Gradients for neural models). Population-level ΔΦj _j is computed via SHAP over the full post-drift held-out set for XGBoost only; audit logs use permutation-based ablation over the 512-sample adaptation subset. The adaptation-induced shift is: ΔΦj=Φj(fadapt,Xpost)−Φj(fsrc,Xpost) _j= _j(f_adapt,X_post)- _j(f_src,X_post) (10) Evaluating both model states on identical post-drift data XpostX_post isolates weight recalibration from distributional change. All ΔΦj _j estimates are over n=5,749n=5,749 post-drift stays with B=1,000B=1,000 bootstrap resamples. 2.6 Attribution-Driven Temporal Retrieval While standard Retrieval-Augmented Generation (RAG) [17] systems synthesize natural language responses, our module adapts this paradigm for predictive retrieval. It avoids static corpora by coupling retrieval to per-instance model attributions and conditions the corpus window on the detected drift era, producing a closed loop between adapted model behaviour and adapted evidence retrieval. At inference, Integrated Gradients [16] are computed over inputs using a zero baseline and 20 steps: ϕj=(xj−xj′)∫01∂F(′+α(−′))∂xjdα _j=(x_j-x _j)\! _0^1\! ∂ F(x +α(x-x ))∂ x_jdα (11) Two sub-queries are built independently: a physiology sub-query from the top-5 |ϕj|| _j| features of X (mask columns excluded) and a treatment sub-query from the top-4 features of Z. Because the physiology LSTM weights are bitwise identical between Run A and Run B by construction, physiology sub-queries remain empirically stable—linking the architectural freeze to retrieval consistency. Queries are prefixed with label anchors (e.g., “septic shock ICU”), encoded via MedCPT [26]—whose asymmetric design (separate encoders for short keyword queries and long article passages) suits the structural mismatch between attribution-derived queries and PubMed abstracts—and matched against an era-conditioned PubMed corpus (871 ICU abstracts retrieved via NCBI E-utilities; 522 retained after restriction to 2009–2019, covering the 10-year window ending the year before detected drift onset). Retrieved sets are merged by PMID, re-ranked by similarity, and the top-k list is returned with the prediction. Retrieval stability is measured as Jaccard overlap of retrieved PMID sets against the source model, assessed on 300 post-drift patients (100 per label). Retrieval quality is evaluated via canonical MeSH descriptors (MeSH P@5, nDCG@5) and a 45-document clinician-rated scaffold (rated by the corresponding author, blinded to model identity; single-rater assessment is a limitation). 2.7 Evaluation Protocol Primary metrics are AUROC and area under the precision-recall curve (AUPRC) per target. Uncertainty is quantified via bootstrap CIs (B=1,000B=1,000): BCa for paired AUROC Δ , percentile for AUPRC Δ [27].Standard 95% bootstrap confidence intervals (α=0.05α=0.05) are reported for the primary AUROC family; AUPRC, Brier scores, and RAG metrics are exploratory. The source model is selected as the median of five-seed ensemble (seeds 42, 123, 7, 2024, 99) on validation AUROC. The two-stream model was trained for up to 50 epochs with early stopping (patience =8=8) using the Adam optimiser (learning rate 1×10−31× 10^-3, weight decay 1×10−51× 10^-5). Adaptation uses Adam (η=3×10−4η=3× 10^-4, up to 40 epochs, patience 8). An XGBoost classifier [24] was trained on a flattened feature representation comprising last-value, mean, standard deviation, minimum, and maximum of each physiological feature across the observation window, concatenated with the 12 treatment features; mask columns are excluded, leaving approximately 54 non-mask sequential features, yielding 5×54+12=2825× 54+12=282 dimensions. Two configurations are evaluated: XGBoost-source (no adaptation) and XGBoost-adapted (retrained on the same post-drift partition used for Run B), serving as our proxy for the industry-standard sliding-window periodic retraining strategy. Post-hoc SHAP (TreeExplainer) analysis [28] was performed on both XGBoost configurations to examine feature attribution stability under drift. No post-hoc probability calibration (Platt scaling or isotonic regression) was applied to any model; all reported probabilities are native model outputs. 3 Results 3.1 Drift Detection and Localisation The dual-gated detector maintained stability through 2014–2019 despite incremental protocol shifts (Sepsis-3 adoption 2016, SSC guideline updates 2016–2018): neither gate was crossed, confirming stable model performance throughout the pre-drift period. The 2020–2022 cohort triggered a severe alert via the distributional leg (composite distributional score 0.2324 vs. training baseline 0.0095); the accuracy leg did not independently fire, though vasopressor’s AUROC drop (0.0197) approached the 0.02 threshold. Five treatment features exceeded drift thresholds: total_crystalloid_ml (PSI=0.76; mean volume fell from 791 to 255 mL), insulin infusion prevalence (−8%-8\%), blood product use (−8%-8\%), PRBC volume, and antibiotic timing—all consistent with documented COVID-era conservative resuscitation [1]. PSI analysis across all 54 physiological features confirmed distributional stability (mean PSI=0.016, max PSI=0.100; 53/54 in the stable range (PSI<<0.10)). This empirically validates the architectural prior: drift was localised predominantly to the treatment domain, with one physiological feature at the minor-drift boundary (PSI = 0.10, 53/54 features stable) (Fig. 1). Figure 1: Distribution shifts in key treatment features: (A) total crystalloid volume and (B) insulin infusion rate. Physiological features (max PSI=0.100) show no comparable shift. 3.2 Pre-Drift Sanity Check and Baseline Equivalence On the pre-drift cohort (2014–2019), Runs A, B, and C are mathematically identical by construction (ΔAUROC=0.0000 =0.0000 across all labels), since adaptation is not yet triggered. Run D uses independently trained source weights to establish its distinct pre-drift baseline. XGBoost-Adapted is identical to XGBoost-Source, as retraining has not occurred. The stable performance across all architectures prior to adaptation rules out initialisation bias: any post-drift divergence is attributable solely to the differing adaptation strategies and architectural priors rather than baseline discrepancies. 3.3 Post-Drift Discrimination and Calibration Post-drift evaluation set: n=5,749n=5,749 stays (12 removed by subject-level deduplication from the nominal 60% of 9,602). Of the 106 septic shock cases in the full post-drift cohort (n=9,602n=9,602; prevalence 1.1%), 76 fell within this held-out evaluation partition (prevalence 1.3%), with the remainder allocated to adaptation training and validation. Results are in Tables 2 and 4 and Fig. 2. Run B (selective adaptation) achieved mAUROC=0.9316, outperforming frozen Run A (0.8965) with the largest gain for vasopressor (Δ=+0.0713 =+0.0713, BCa 95% CI [+0.0622,+0.0822][+0.0622,+0.0822], p<0.001p<0.001) and a significant gain for septic shock (Δ=+0.0303 =+0.0303, BCa 95% CI [+0.0113,+0.0497][+0.0113,+0.0497], p<0.001p<0.001). Run B surpassed both unconstrained full adaptation (Run C, 0.9249) and single-stream last-layer fine-tuning (Run D, 0.9010), confirming that structural decomposition—not merely selective freezing—is the operative mechanism. XGBoost-adapted achieved marginally higher mean AUROC (0.9382), but this aggregate advantage has hidden bedside costs detailed in Section 3.4. Per-label bootstrap 95% confidence intervals for all four neural configurations are reported in Table 3. Mean AUROC is a necessary but not sufficient criterion for model selection in high-stakes clinical settings [32]. AUPRC results are decisive for the rare septic shock condition (1.3% prevalence): Run B improved from 0.3100 (Run A) to 0.4131 (+0.1031+0.1031, 95% CI [+0.0484,+0.1553][+0.0484,+0.1553]). Bootstrap 95% CIs on all paired AUPRC gains confirm statistical reliability: vasopressor [+0.1721,+0.2301][+0.1721,+0.2301], intubation [+0.0007,+0.0159][+0.0007,+0.0159], septic shock [+0.0484,+0.1553][+0.0484,+0.1553]. The vasopressor gain is substantial and precisely estimated; the septic shock CI reflects the small positive class (n=76n=76); the modest intubation gain is consistent with the absence of ventilation-adjacent treatment features. By comparison, Run D deteriorated septic shock AUPRC to 0.2149, further demonstrating that entangled adaptation damages rare-class performance. XGBoost-adapted’s septic shock AUPRC degraded from 0.3313 to 0.2600 after retraining—evidence of probability mass compression toward the majority class, indicating monolithic retraining actively traded precision-recall performance on the rarest, most dangerous class in exchange for aggregate discrimination. Run B restored calibration across all targets (vasopressor Brier 0.1861→0.12410.1861→ 0.1241; intubation 0.1223→0.10160.1223→ 0.1016; septic shock 0.0613→0.01840.0613→ 0.0184). XGBoost-adapted’s nominal Brier score (septic shock 0.0116) reflects probability compression rather than genuine calibration—a model assigning near-zero probabilities to all patients minimises squared error under severe class imbalance without detecting any positives at a meaningful threshold, confirmed threshold-free by the AUPRC collapse from 0.3313 to 0.2600. Figure 2: Post-drift AUROC trajectories across three labels (vasopressor, intubation, septic shock). Runs A/B/C are identical on pre-drift data; adaptation diverges only post-2020. Run B (selective) outperforms both frozen Run A and single-stream Run D. Table 2: Post-Drift AUROC and Mean AUROC (n=5,749n=5,749). Model Vaso Intub Shock Mean Monolithic Baselines XGBoost (Source) 0.9148 0.9475 0.9367 0.9330 XGBoost (Adapted) 0.9330 0.9413 0.9404 0.9382 Neural Baselines Run A: Static Source 0.8233 0.9367 0.9296 0.8965 Run C: Full Adapt 0.8779 0.9435 0.9533 0.9249 Run D: Single-Stream 0.8269 0.9414 0.9347 0.9010 Proposed Framework Run B: Selective Adapt 0.8947 0.9403 0.9599 0.9316 Table 3: Post-Drift AUROC with 95% Bootstrap Confidence Intervals (percentile method, n=5,749n=5,749). XGBoost point estimates are in Table 2; bootstrap CIs were not computed for it. Label Run A Run B Run C Run D Vasopressor 0.823 [0.805, 0.841] 0.895 [0.879, 0.910] 0.878 [0.862, 0.893] 0.827 [0.810, 0.843] Intubation 0.937 [0.930, 0.943] 0.940 [0.934, 0.947] 0.943 [0.937, 0.950] 0.941 [0.935, 0.948] Septic Shock 0.930 [0.904, 0.953] 0.960 [0.934, 0.978] 0.953 [0.925, 0.975] 0.935 [0.907, 0.957] Table 4: Post-Drift AUPRC and Septic Shock Brier (n=5,749n=5,749). Model Vaso Intub Shock Brier (↓ ) Monolithic Baselines XGBoost (Source) 0.6487 0.8757 0.3313 0.0123 XGBoost (Adapted) 0.6612 0.8586 0.2600 0.0116 Neural Baselines Run A: Static Source 0.4187 0.8475 0.3100 0.0613 Run C: Full Adapt 0.5714 0.8632 0.3775 0.0167 Run D: Single-Stream 0.3544 0.8598 0.2149 0.0324 Proposed Framework Run B: Selective Adapt 0.6182 0.8558 0.4131 0.0184 3.4 Clinical Safety and Bedside Disagreement The aggregate AUROC advantage of XGBoost-adapted is accompanied by a severe unidirectional failure pattern. Defining a “Catch” as p≥0.50p≥ 0.50 and a “Miss” as p<0.10p<0.10 on a ground-truth positive, Table 5 shows that Run B caught 26 true-positive septic shock cases that XGBoost-adapted critically missed, with zero cases in the reverse direction. This 26/0 asymmetry persists across thresholds (31/0 at catch ≥0.40≥ 0.40; 32/0 at catch ≥0.30≥ 0.30; 19/0 at miss <0.05<0.05) and is confirmed threshold-free by the AUPRC collapse. Run C (20/0) and Run D (25/0) share this directional advantage, confirming monolithic retraining suppresses septic shock sensitivity across all comparators; Table 6 reports the full confusion matrix at τ=0.5τ=0.5 for both models on all three targets. Notably, XGBoost’s own septic-shock performance degrades on both axes under adaptation—not merely a precision-recall tradeoff—dropping from 34.2% recall / 40.0% precision (source) to 6.6% recall / 35.7% precision (adapted), indicating that retraining does not sharpen the model but rather suppresses it into near-silence on the rarest, most lethal target. Relative to XGBoost-adapted, Run B raises septic-shock recall from 6.6% to 69.7% (5 vs. 53 of 76 true cases caught) at a cost of 115 additional false positives out of 5,673 total negatives (an absolute false-alarm-rate increase of 2.0 percentage points), or approximately 2.4 extra alarms per additional true case caught. This tradeoff is markedly less favourable for vasopressor (6.9 extra alarms per additional catch) and intubation (5.1), indicating that recall-prioritisation is best justified specifically for the rarest and most lethal target rather than as a blanket property of the architecture. This failure is illustrated by specific patient trajectories. For vasopressor prediction (Patient A, stay_id 35773744), Run B predicted 90.4% vs. XGBoost’s 15.5%. For septic shock (Patient B, stay_id 31656477), Run B assigned 85.6% probability; XGBoost assigned 4.0%—a critically low probability for a confirmed lethal condition. These cases confirm that monolithic retraining suppresses true-positive alerts despite acceptable aggregate metrics. Table 5: Bedside Disagreement (catches/misses vs. XGBoost-Adapted, n=5,749n=5,749). Catch: p≥0.50p≥ 0.50; Miss: p<0.10p<0.10. Model Vaso (n=554n=554) Intub (n=1657n=1657) Shock (n=76n=76) Run B: Selective Adapt 19 / 0 4 / 0 26 / 0 Run C: Full Adapt 17 / 0 7 / 0 20 / 0 Run D: Single-Stream 14 / 0 6 / 0 25 / 0 Table 6: XGBoost-Adapted vs. Run B at τ=0.5τ=0.5 (TP/FP/FN and Sensitivity/Precision). Target Model TP FP FN Recall Prec. Septic Shock (n=76n=76) XGBoost (adapted) 5 9 71 6.6% 35.7% Run B (proposed) 53 124 23 69.7% 29.9% Vasopressor (n=554n=554) XGBoost (adapted) 355 239 199 64.1% 59.8% Run B (proposed) 441 836 113 79.6% 34.5% Intubation (n=1,657n=1,657) XGBoost (adapted) 1428 368 229 86.2% 79.5% Run B (proposed) 1464 553 193 88.4% 72.6% 3.5 Biological Amnesia via Δ -Attribution The Two-Stream architecture provides a mathematically exact guarantee: LSTM physiology parameters are bitwise identical between Run A and Run B (physio mean_rel_Δ=0.0000 =0.0000). Any attribution shift in Run B originates exclusively from the updated fusion head re-weighting a structurally preserved biological state—a property unreplicable by post-hoc methods on monolithic models [29]. The following ΔΦ analysis is self-contained within XGBoost (SHAP values); attribution magnitudes are not compared across architectures as SHAP and Integrated Gradients operate on incomparable scales. In XGBoost-adapted, ΔΦ analysis reveals treatment shifts that are clinically coherent (mean |SHAP||SHAP| for total_crystalloid_ml: 1.057→2.1851.057→ 2.185, ΔΦ=+1.128 =+1.128 for vasopressor), accompanied by simultaneous recalibration of stable physiological features (PSI<0.10<0.10): max_lactate lost ΔΦ=−0.438[−0.442,−0.434]∗ =-0.438\,[-0.442,-0.434]^*; std_ph_venous gained +0.847[+0.835,+0.859]∗+0.847\,[+0.835,+0.859]^*—a near-inversion of the canonical haemodynamic severity signal (∗95% CI excludes zero). This is the biological amnesia introduced above: the measurable, population-level overwriting of stable physiological representations in a monolithic model (Fig. 3). Population-level Δ -Attribution analysis identified statistically significant importance shifts across 93.6% of features (264 of 282) for vasopressor (95.7% for septic shock, 92.9% for intubation); for vasopressor, gains outnumbered losses among physiological features (160 vs. 93), while for septic shock the pattern reversed, with losses dominating (226 vs. 32), confirming widespread physiological recalibration in both directions across labels. Figure 3: Biological amnesia: illustrative single-patient SHAP comparison pre/post-adaptation. XGBoost-adapted shows Δ=1.454 =1.454 SHAP units on total_crystalloid_ml with widespread physiological recalibration; the Two-Stream framework shows zero physiological shift. 3.6 Attribution-Driven Temporal Retrieval The frozen physiology LSTM structurally anchors evidence retrieval. Table 7 reports Jaccard overlap of retrieved PMID sets against the source model across 300 post-drift patients (100 per label). Run B maintains physiology-stream Jaccard 0.573 vs. 0.330 for XGBoost-adapted (Δ=+0.243 =+0.243); the treatment stream adapts as expected (0.540 vs. 0.448). Per-label physiology Jaccard: vasopressor 0.485/0.260, intubation 0.619/0.293, septic shock 0.614/0.438. Query-level token Jaccard confirms the structural cause: Run B physiology queries share 72.9% of tokens with the source model vs. 34.1% for XGBoost—the direct consequence of identical LSTM weights between Run A and Run B. Table 7: Evidence Retrieval Stability (n=300n=300 post-drift patients). Jaccard overlap of retrieved PubMed documents vs. source model; higher is better. Metric Run B XGBoost (Adp.) Physiology Jaccard (↑ ) 0.573 0.330 Treatment Jaccard (↑ ) 0.540 0.448 Merged Jaccard (↑ ) 0.438 0.323 Rank corr., physio Spearman (↑ ) 0.465 0.131 Rank corr., treatment Spearman (↑ ) 0.592 0.439 Token Jaccard, physio (↑ ) 0.729 0.341 Token Jaccard, treatment (↑ ) 0.717 0.602 The mechanistic link between architectural freeze and retrieval stability is confirmed via Spearman correlation between per-patient Δphysio _physio and Jaccard divergence: for Run B, r=−0.197r=-0.197 (p=0.0006p=0.0006, computed on n=241n=241 cases); for XGBoost-adapted, r=−0.169r=-0.169 (p=0.0033p=0.0033, computed on only n=137n=137 cases, as the remaining 163 cases were completely excluded due to zero document overlap with the source queries). The negative sign for Run B reflects that its residual attribution delta originates from the fusion head—not physiological features—so retrieval remains anchored to source documents regardless of fusion-head recalibration magnitude. Automatic retrieval quality (Table 8) shows equivalent MeSH P@5 across all configurations (0.635/0.635/0.632), confirming selective adaptation does not degrade topical relevance; per-label, vasopressor achieves the highest precision (P@5 0.710–0.760) and intubation the lowest (P@5 0.208–0.252), reflecting the broader MeSH vocabulary of mechanical ventilation literature. Clinician-rated P@5 favours Run B (0.800 vs. XGBoost 0.467)—a gap substantially wider than the automatic MeSH P@5 gap (0.635 vs. 0.632)—while XGBoost’s higher automatic nDCG@5 (0.913 vs. 0.838, Table 8) reflects superior ranking of the relevant documents it retrieves, rather than broader topical coverage. Table 8: Automatic Retrieval Quality. MeSH P@5 and nDCG@5 over 300 post-drift cases using canonical MeSH descriptors as oracle. Label / Overall Source Run B XGBoost (Adp.) P@5 Overall 0.635 0.635 0.632 P@5 Vasopressor 0.750 0.760 0.710 P@5 Intubation 0.252 0.248 0.208 P@5 Septic Shock 0.902 0.896 0.978 nDCG@5 Overall 0.860 0.838 0.913 Worked examples. The clinical value is best understood at the individual patient level. We present three representative cases drawn from the RAG evaluation cohort– including a mechanistic deep-dive into the vasopressor disagreement case introduced in Section 3.4—to illustrate the clinical value at the individual patient level. These cases capture both rescue scenarios—where the baseline model critically misses the diagnosis—and spurious reasoning, where the baseline model predicts correctly but for physiologically incoherent reasons. Patient A, vasopressor (stay 35773744). True label =1=1: Run B assigned probability 0.904; XGBoost-adapted assigned 0.155—a confident false negative. Integrated Gradients identified lactate, lactate_baseline, and diastolic blood pressure as the dominant physiology-stream contributors. The physiology sub-query retrieved haemodynamic management abstracts with identical PMID sets for source and Run B (Jaccard =1.000=1.000); the XGBoost query—shifted toward fluid and medication volume terms by retraining drift—retrieved a substantially different and less clinically relevant document set. Patient B, intubation (stay 31841598). Both Run B (0.926) and XGBoost (0.960) correctly predicted mechanical ventilation. Integrated Gradients identified gcs_verbal, gcs_eye, and respiratory rate as the dominant physiology contributors. Source and Run B physiology queries retrieved identical documents (Jaccard =1.000=1.000); crucially, XGBoost’s internal query—incorporating WBC, age, and antibiotic timing from the flattened feature representation—retrieved a substantially different set skewed toward antibiotic management literature. This illustrates spurious reasoning: a correct prediction grounded in the wrong physiological evidence, exposing attribution redistribution despite an accurate outcome. Patient C, septic shock (stay 30855786). Run B assigned probability 0.805; XGBoost-adapted assigned 0.039—missing a highly lethal condition. Integrated Gradients identified diastolic blood pressure, lactate_baseline, and urine output as the dominant physiology contributors. The Run B physiology sub-query retrieved the same documents as the source model (Jaccard =1.000=1.000), including abstracts on cardiovascular determinants of sepsis resuscitation; the XGBoost query—shifted toward respiratory and metabolic terms by attribution redistribution under retraining—retrieved a divergent set. The frozen physiology stream thus provides two complementary benefits: preserving both the predictive signal and the evidence grounding for the physiological features defining each patient’s clinical presentation. 3.7 Sensitivity Analysis and Ablation Studies To validate robustness, we re-ran Run B selective adaptation varying the post-drift training split ratio (20/30/40%, fixed 500-stay replay buffer) and the replay buffer size (0/250/500 stays, fixed 30% split). The 30% ratio yielded the strongest performance among the three tested values. Across replay buffer sizes, post-drift discrimination was essentially unchanged (mean AUROC within 0.001–0.002 of one another), confirming that the frozen physiology stream—not the replay buffer—is the primary anti-forgetting mechanism. The 500-stay buffer achieved strongest pre-drift retention among the three tested, serving a complementary backward-compatibility role. 4 Discussion Results suggest that CDSS temporal degradation is driven predominantly by treatment-protocol drift rather than physiological change—a distinction current monolithic adaptation methods cannot enforce structurally. The two-stream architecture exploits this asymmetry by design: freezing the physiology LSTM is not a heuristic choice but an empirically justified structural prior (treatment max PSI 0.76 vs. physiological max PSI 0.10). The result is a mathematically exact guarantee that no post-hoc explanation method can provide: physiological representations are bitwise identical between source and adapted model, so any attribution shift originates solely from the updated fusion head. The 26/0 unidirectional septic shock disagreement is the most clinically striking result. XGBoost-adapted achieved marginally higher mean AUROC, but this aggregate advantage actively suppressed true-positive alerts on the rarest and most lethal condition—trading precision-recall performance for aggregate rank-ordering. The Δ -Attribution analysis explains why: monolithic retraining simultaneously recalibrated up to 95.7% of features (label-dependent), including stable physiological features whose PSI confirmed no distributional change, producing shifts that are architecturally impossible to distinguish from legitimate co-adaptation. This is biological amnesia as a measurable, population-level phenomenon. Prior work (Nestor et al. [7], Futoma et al. [10]) characterised the problem; this paper provides both an architectural remedy and a formal metric to detect and audit it. The attribution-driven Temporal RAG closes a complementary gap: explanatory staleness. Unlike Almanac [30] and similar static-corpus systems, our module couples retrieval to per-instance attributions and to the publication era of the source training window, ensuring a 2022 prediction is grounded in contemporaneous evidence (2009–2019) rather than potentially divergent post-pandemic guidelines. Clinical implications. Drift detection confirmed that treatment-side shift can be clinically substantial even when patient physiology remains stable—the five drifted features (conservative crystalloid resuscitation PSI=0.76, reduced insulin infusion −8%-8\%, reduced blood product use −8%-8\%, PRBC volume, accelerated antibiotic timing) are all consistent with documented COVID-era protocol changes [1]. This supports recent calls for ongoing post-deployment performance monitoring rather than one-time validation [12, 31]. Calibration improvements are equally significant: selective adaptation recalibrates probability estimates, not merely ranks (vasopressor Brier 0.1861→0.12410.1861→ 0.1241; intubation 0.1223→0.10160.1223→ 0.1016; septic shock 0.0613→0.01840.0613→ 0.0184). Since clinical intervention thresholds are probability-based, CDSS deployed against miscalibrated outputs can produce systematically biased recommendations even at high AUROC [32]. The attribution audit logs address this governance requirement directly: by documenting per-feature importance shifts at each adaptation event, the framework makes model updates interpretable and contestable by clinicians—a property monolithic retraining cannot provide by design. Relation to prior work. Prior work on temporal generalisation in clinical ML has predominantly characterised the problem rather than offered architectural remedies. Nestor et al. [7] demonstrated feature-importance instability across MIMIC-I eras but did not propose adaptation; Futoma et al. [10] argued that single-split evaluation systematically overstates clinical utility. Compared to continual-learning regularisation (EWC [11]), which imposes a soft hyperparameter-dependent penalty, our structural freeze is unconditional: a physiologically stable feature with small source-domain weights is not protected by EWC but is fully protected by our architecture. Domain-adversarial approaches [33] require environmental labels during source training and do not explicitly isolate treatment shifts; head-to-head comparison is a planned extension. On the explanation side, recent clinical RAG systems such as Almanac [30] rely on static corpora; our framework couples retrieval to per-instance attributions and to publication-era filtering, producing patient-specific and time-appropriate evidence in a single closed loop. The Δ -Attribution metric extends concept-drift detection via model explanation [25] to a formally defined, population-level measure of physiological recalibration under adaptation—a distinction Demšar and Bosnić did not address. Together, these elements constitute a pipeline in which architectural design, adaptation governance, and evidence retrieval are structurally coupled rather than independently applied—a combination not present in prior clinical ML literature to our knowledge. Limitations. Evaluation is limited to a single institution; external validation on eICU-CRD is the natural next step. The treatment vector lacks non-leaking ventilation-adjacent features, limiting intubation-specific adaptation benefit; identifying such proxies is a natural extension. Septic shock results (76 positives) warrant replication on larger post-2020 cohorts. The RAG module currently surfaces abstract-level evidence; production use would benefit from full-text retrieval with LLM summarisation. The RAG evaluation relies on a single-rater clinician scaffold; multi-rater validation would strengthen the reliability of the P@5 and nDCG@5 estimates. Ethnicity indicators in the treatment vector carry equity risks requiring fairness evaluation. Head-to-head benchmarking against EWC and DANN is left to future work. All analysis is retrospective; prospective impact measurement is needed [8]. 5 Conclusion We have presented a governable clinical intelligence architecture that moves the CDSS paradigm from static, monolithic models toward dynamic, attribution-aware systems. By structurally decoupling stable human physiology from evolving institutional treatment protocols, our framework overcomes the biological amnesia inherent in standard monolithic retraining. We demonstrated that clinical AI can evolve safely—not as an opaque update, but as a transparent, governed process grounded in causal audit logs and per-instance evidence retrieval. Our results validate this approach on the MIMIC-IV cohort, where selective adaptation achieved superior bedside safety—most notably catching 26 critical septic shock cases missed by standard retraining—while maintaining retrieval consistency as the model evolved. Ultimately, this framework enables CDSS to keep pace with clinical practice while preserving the integrity of fundamental patient biology. Ethics and Data MIMIC-IV (v3.1) is de-identified and publicly available via PhysioNet (https://physionet.org/content/mimiciv/3.1/); IRB approval was not required. Code and the curated PubMed corpus will be released at the corresponding author’s GitHub upon acceptance. The authors declare no competing interests and received no specific funding. Acknowledgment During the preparation of this work, the authors used Claude (Anthropic), Grok (xAI), and Gemini (Google) for language editing, LaTeX formatting, and structural organisation of the manuscript text. No AI system was used to generate, analyse, or interpret the research data or results. The authors reviewed and edited all AI-assisted content and take full responsibility for the accuracy and integrity of the published work. References [1] L. Evans et al., “Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021,” Crit. Care Med., vol. 49, no. 11, p. e1063–e1143, 2021, doi: 10.1097/CCM.0000000000005337. [2] A. Rhodes et al., “Surviving sepsis campaign: international guidelines for management of sepsis and septic shock: 2016,” Crit. Care Med., vol. 45, no. 3, p. 486–552, 2017, doi: 10.1097/CCM.0000000000002255. [3] J. W. Devlin et al., “Clinical practice guidelines for pain, agitation/sedation, delirium, immobility, and sleep in ICU patients,” Crit. Care Med., vol. 46, no. 9, p. e825–e873, 2018, doi: 10.1097/CCM.0000000000003299. [4] G. Grasselli et al., “Baseline characteristics and outcomes of 1591 patients infected with SARS-CoV-2 admitted to ICUs of the Lombardy region, Italy,” JAMA, vol. 323, no. 16, p. 1574–1581, 2020, doi: 10.1001/jama.2020.5394. [5] J. Gama et al., “A survey on concept drift adaptation,” ACM Comput. Surveys, vol. 46, no. 4, Art. no. 44, 2014, doi: 10.1145/2523813. [6] J. Lu et al., “Learning under concept drift: a review,” IEEE Trans. Knowl. Data Eng., vol. 31, no. 12, p. 2346–2363, 2019, doi: 10.1109/TKDE.2018.2876857. [7] B. Nestor et al., “Feature robustness in non-stationary health records: caveats to deployable model performance in common clinical machine learning tasks,” in Proc. MLHC, vol. 106, 2019, p. 381–405. [8] C. J. Kelly et al., “Key challenges for delivering clinical impact with artificial intelligence,” BMC Med., vol. 17, no. 1, Art. no. 195, 2019, doi: 10.1186/s12916-019-1382-x. [9] A. Wong et al., “External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients,” JAMA Intern. Med., vol. 181, no. 8, p. 1065–1070, 2021, doi: 10.1001/jamainternmed.2021.2626. [10] J. Futoma et al., “The myth of generalisability in clinical research and machine learning in health care,” Lancet Digit. Health, vol. 2, no. 9, p. e489–e492, 2020, doi: 10.1016/S2589-7500(20)30186-2. [11] J. Kirkpatrick et al., “Overcoming catastrophic forgetting in neural networks,” PNAS, vol. 114, no. 13, p. 3521–3526, 2017, doi: 10.1073/pnas.1611835114. [12] J. Feng et al., “Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare,” npj Digit. Med., vol. 5, no. 1, Art. no. 66, 2022, doi: 10.1038/s41746-022-00611-y. [13] M. B. McDermott et al., “Reproducibility in machine learning for health research: still a ways to go,” Sci. Transl. Med., vol. 13, no. 586, Art. no. eabb1655, 2021, doi: 10.1126/scitranslmed.abb1655. [14] H. Zhang et al., “Shifting machine learning for healthcare from development to deployment and from models to data,” npj Digit. Med., vol. 5, no. 1, Art. no. 40, 2022, doi: 10.1038/s41746-022-00698-3. [15] V. Lösing, B. Hammer, and H. Wersing, “Incremental on-line learning: a review and comparison of state of the art algorithms,” Neurocomputing, vol. 275, p. 1261–1274, 2018, doi: 10.1016/j.neucom.2017.06.084. [16] M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” in ICML, 2017, p. 3319–3328. [17] P. Lewis et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in NeurIPS, vol. 33, 2020, p. 9459–9474. [18] A. E. Johnson et al., “MIMIC-IV, a freely accessible electronic health record dataset,” Sci. Data, vol. 10, no. 1, Art. no. 1, 2023, doi: 10.1038/s41597-022-01899-x. [19] M. Moor et al., “Early prediction of sepsis in the ICU without target leakage: a new benchmark,” Front. Med., vol. 8, Art. no. 607952, 2021, doi: 10.3389/fmed.2021.607952. [20] M. Singer et al., “The third international consensus definitions for sepsis and septic shock (Sepsis-3),” JAMA, vol. 315, no. 8, p. 801–810, 2016, doi: 10.1001/jama.2016.0287. [21] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Comput., vol. 9, no. 8, p. 1735–1780, 1997, doi: 10.1162/neco.1997.9.8.1735. [22] J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv:1607.06450 [stat.ML], 2016. [23] T.-Y. Lin et al., “Focal loss for dense object detection,” in ICCV, 2017, p. 2980–2988, doi: 10.1109/ICCV.2017.324. [24] T. Chen and C. Guestrin, “XGBoost: a scalable tree boosting system,” in KDD, 2016, p. 785–794, doi: 10.1145/2939672.2939785. [25] J. Demšar and Z. Bosnić, “Detecting concept drift in data streams using model explanation,” Expert Syst. Appl., vol. 92, p. 546–559, 2018, doi: 10.1016/j.eswa.2017.10.003. [26] Q. Jin et al., “MedCPT: contrastive pre-trained transformers with large-scale PubMed search logs for zero-shot biomedical information retrieval,” Bioinformatics, vol. 39, no. 11, Art. no. btad651, 2023, doi: 10.1093/bioinformatics/btad651. [27] B. Efron, “Better bootstrap confidence intervals,” J. Amer. Stat. Assoc., vol. 82, no. 397, p. 171–185, 1987, doi: 10.1080/01621459.1987.10478410. [28] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in NeurIPS, vol. 30, 2017, p. 4765–4774. [29] C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,” Nature Mach. Intell., vol. 1, no. 5, p. 206–215, 2019, doi: 10.1038/s42256-019-0048-x. [30] C. Zakka et al., “Almanac—retrieval-augmented language models for clinical medicine,” NEJM AI, vol. 1, no. 2, Art. no. AIoa2300068, 2024, doi: 10.1056/AIoa2300068. [31] S. G. Finlayson et al., “The clinician and dataset shift in artificial intelligence,” N. Engl. J. Med., vol. 385, no. 3, p. 283–286, 2021, doi: 10.1056/NEJMc2104626. [32] B. Van Calster et al., “Calibration: the Achilles heel of predictive analytics,” BMC Med., vol. 17, no. 1, Art. no. 230, 2019, doi: 10.1186/s12916-019-1466-7. [33] Y. Ganin et al., “Domain-adversarial training of neural networks,” J. Mach. Learn. Res., vol. 17, no. 59, p. 2096–2130, 2016.