Odds-Structured Distributional Regression for Interpretable Time-to-Event Prediction
Authors: Jakob Versnjak, Manuel Welte, Titus Kuehne, Marcus Kelm, Grégoire Montavon
Organizations: BIFOLD – Berlin Institute for the Foundations of Learning and Data, Berlin, Germany · Deutsches Herzzentrum der Charité, Institute of Computer-assisted Cardiovascular Medicine, Augustenburger Platz 1, 13353 Berlin, Germany · Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Charitéplatz 1, Berlin 10117, Germany · Deutsches Herzzentrum der Charité, Institute for AI in Medicine, Invalidenstraße 120/121, 10115 Berlin, Germany · Department of Congenital Heart Disease – Pediatric Cardiology, Deutsches Herzzentrum der Charité, Augustenburger Platz 1, Berlin 13353, Germany · DZHK (German Centre for Cardiovascular Research), Partner Site Berlin, Potsdamer Str. 58, Berlin 10785, Germany
Time-to-event prediction is fundamentally challenged by partial observability and aleatoric uncertainty. In clinical medicine, for instance, the timing of disease progression or acute readmission is contingent on highly dynamic and largely latent health trajectories. Rather than forcing a deterministic point forecast, distributional regression offers a principled means to address this complexity by predicting a probability distribution over time. However, explaining these distributional predictions presents a key obstacle: raw distribution parameters output by the model are abstract mathematical quantities that do not align well with domain concepts. To gain interpretability, we propose a novel paradigm where output parameters are designed to be additive in log-odds---a widely established measure in clinical risk modeling. Our approach, called ODDS-DR, moves beyond technical convenience to deliver clear domain alignment as well as the layer-wise computational structure required for robust, time-dependent feature attribution. Validated across two real-world clinical cohorts, our framework matches state-of-the-art predictive performance while resolving distributional outputs into interpretable risk markers across flexible prognostic horizons.
Figures & tables
Figure 1: Time-to-event modeling and proposed ODDS-DR framework. (A) Clinical data records both the occurrence of events and their temporal organization. (B) Distributional regression models this temporal dynamic by parameterizing continuous, patient-specific probability densities over time. (C) Classical baseline: Standard distributional regression predicts distribution-specific parameters θ(x) (e.g., shape and scale) that lack clinical interpretability and decomposability into input features. (D) ODDS-DR (ours): By parameterizing model representations to contribute linearly to event log-odds L(x,t) —a clinically established quantity—our framework enables exact linear attribution to intermediate representations, followed by propagation (e.g., via LRP) back to input clinical features.
Figure 2: Comparison of parameterization geometries for a Gaussian distribution. Left (Standard parameterization): Classical formulation with location μ=zμ and scale σ=softplus(zσ) , where variance changes monotonically along the vertical axis independently of location. Right (ODDS-DR, log-odds-additive parameterization): Our reparameterization (z1,z2) places maximum uncertainty (spread) at the origin (0,0) , with lateral displacement along z1 scaling proportionally with density spread. This geometry guarantees exact log-odds additivity and facilitates feature attribution.
Figure 3: ODDS-DR behavior on synthetic time-to-event data with heterogeneous risk profiles and representativeness gaps. Top: Observed and censored event times across four subpopulations: general population, under-represented, risk group A (events concentrated early at t≤2 ), and risk group B (events distributed over an extended horizon t≤10 ). Bottom left: Predicted log-odds across subpopulations evaluated at horizons t∈{2,6,10} , decomposing total risk ( λ1z1+λ2z2 , pink) into time-invariant baseline risk ( λ1z1 , dashed) and time-dependent adjustment. Bottom right: Continuous risk trajectories over time for risk groups A and B, demonstrating how ODDS-DR captures distinct temporal risk dynamics.
Model
Distribution
SUPPORT2
CoA
ODDS-DR ( ours )
Gaussian
0.80 (0.79–0.81)
0.66 (0.52–0.82)
ODDS-DR ( ours )
Logistic
0.81 (0.80–0.82)
0.60 (0.44–0.77)
ODDS-DR ( ours )
Log-logistic
0.83 (0.82–0.83)
0.64 (0.50–0.77)
Ablation study:
no reparameterization
Gaussian
0.81 (0.80–0.82)
0.61 (0.44–0.77)
no reparameterization
Logistic
0.80 (0.79–0.81)
0.61 (0.43–0.80)
Table 2: Discriminative performance (Harrell’s C-index) on the SUPPORT2 and CoA validation sets across different distribution choices for ODDS-DR, an ablation study using standard parameterizations (no reparameterization), and competing baseline models. For C-index computation, distributional models and DeepHit were ranked by the restricted area under the predicted survival curve; DeepSurv, CoxNAM, and Cox by the log-risk scores, and Cox-Time by the cumulative hazard evaluated at the median observed event time of the training set. Values in parentheses report 95% confidence intervals obtained via non-parametric bootstrapping with 100 resamples. Bold denotes the best performance; underlined values indicate the second best.
Figure 4: Time-varying feature attribution across prediction horizons. Results are shown for (A) SUPPORT2 and (B) CoA cohorts across increasing follow-up times. Denoting by Rit(n) the attribution of log-odds L onto feature xi at horizon t for patient n , the left panels display mean absolute attribution scores, Rit=En[∣Rit(n)∣] , reflecting overall feature importance. The right panels display feature-wise normalized scores, Rit/maxt′Rit′ , to isolate and compare relative temporal trajectories across clinical markers.
Clinical decisions are high-stakes and require explicit justification, making model interpretability essential for auditing deep clinical models prior to deployment. As the ecosystem of model architectures and explainability methods expands, critical questions remain: Do architectural features like attention improve explainability? Do interpretability approaches generalize across clinical tasks? While prior benchmarking efforts exist, they often lack extensibility and reproducibility, and critically, fail to systematically examine how interpretability varies across the interplay of clinical tasks and model architectures. To address these gaps, we present a comprehensive benchmark evaluating interpretability methods across diverse clinical prediction tasks and model architectures. Our analysis reveals that: (1) attention when leveraged properly is a highly efficient approach for faithfully interpreting model predictions; (2) black-box interpreters like KernelSHAP and LIME are computationally infeasible for time-series clinical prediction tasks; and (3) several interpretability approaches are too unreliable to be trustworthy. From our findings, we discuss several guidelines on improving interpretability within clinical predictive pipelines. To support reproducibility and extensibility, we provide our implementations via PyHealth, a well-documented open-source framework: https://github.com/sunlabuiuc/PyHealth.
Yongda Fan, John Wu, Andrea Fitzpatrick +3
University of Illinois Urbana-Champaign, Champaign, IL 61801, USA · PyHealth · University of Illinois College of Medicine, Chicago, IL 60612, USA
Electronic Health Record (EHR) prediction models in the intensive care unit must learn from sparse and irregular measurements while preserving the clinical meaning of time and supporting transparent decision-making. We present INTERVenE, a family of Transformer architectures whose input is an interval-based, knowledge-based temporal abstraction (KBTA), a token stream of named clinical concepts (states, trends, events, contexts) drawn from a curated medical ontology, rather than an unnamed bin index or a raw measurement triplet. This naming layer is what we ask KBTA to do: it makes the model's per-token attributions resolve to clinical concepts by construction. INTERVenE offers two complementary variants: an auto-regressive decoder that generates future abstraction trajectories with a per-step risk readout (localizing when and after which events risk rises), and a bidirectional encoder for single-pass joint risk and time-to-event prediction. On 57,078 MIMIC-IV admissions, INTERVenE-Enc outperformed GRU-D, STraTS, and KarmaLego, reaching a support-weighted AUPRC of 0.672 (+0.046 over the strongest neural baseline, with non-overlapping 95% bootstrap CIs), AUROC of 0.901, and length-of-stay MAE of 44.4,h. INTERVenE-Ar complements these predictions with generated abstraction trajectories and evolving risk scores, achieving AUROC of 0.854 and AUPRC of 0.588 under the harder generative readout. An alternative-discretization ablation preserved predictive performance, supporting KBTA's role as a clinically named attribution substrate rather than a prerequisite for accuracy.
Shahar Oded, Yuval Shahar
The Stein Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Beer-Sheva, Israel
Accurate and interpretable mortality risk prediction in intensive care units (ICUs) remains a critical challenge due to the irregular temporal structure of electronic health records (EHRs), the complexity of longitudinal disease trajectories, and the lack of clinically grounded explanations in many data-driven models. To address these challenges, we propose TA-RNN-Medical-Hybrid, a time-aware and knowledge-enriched deep learning framework that jointly models longitudinal clinical sequences and irregular temporal dynamics through explicit continuous-time encoding, along with standardized medical concept representations. The proposed framework extends time-aware recurrent modeling by integrating explicit continuous-time embeddings that operate independently of visit indexing, SNOMED-based disease representations, and a hierarchical dual-level attention mechanism that captures both visit-level temporal importance and feature/concept-level clinical relevance. This design enables accurate mortality risk estimation while providing transparent and clinically meaningful explanations aligned with established medical knowledge. We evaluate the proposed approach on the MIMIC-III critical care dataset and compare it against strong time-aware and sequential baselines. Experimental results demonstrate that TA-RNN-Medical-Hybrid consistently improves predictive performance in terms of AUC, accuracy, and recall-oriented F2-score. Moreover, qualitative analysis shows that the model effectively decomposes mortality risk across time and clinical concepts, yielding interpretable insights into disease severity, chronicity, and temporal progression. Overall, the proposed framework bridges the gap between predictive accuracy and clinical interpretability, offering a scalable and transparent solution for high-stakes ICU decision support systems.
Department of Computer Engineering, SRC.,Islamic Azad university, Tehran, Iran · School of Computer Engineering, Iran university of Science and technology, Tehran, Iran