STDP-inspired temporal transition modeling for adaptive clinical risk prediction from electronic health records

Technical Summary: STDP-Inspired Temporal Transition Modeling for Adaptive Clinical Risk Prediction

Problem Statement
Current clinical prediction models utilizing Electronic Health Records (EHR) frequently aggregate longitudinal patient histories into static, patient-level features (e.g., cumulative comorbidity counts or raw lab values). This approach often discards the temporal ordering of clinical events, which can be critical for understanding pathophysiology. For instance, two patients with identical sets of diagnoses and lab abnormalities may have vastly different prognoses if one experienced a rapid, worsening sequence of events while the other accumulated them over a stable period. While temporal modeling techniques like recurrent neural networks (RNNs) and transformers exist, they often require large datasets, significant computational resources, and produce latent representations that lack clinical interpretability. Furthermore, many temporal models are evaluated as endpoint predictors rather than as rolling decision tools, potentially conflating early warning signals with proximal clinical deterioration.

Methodology
The authors propose a simplified framework inspired by Spike-Timing-Dependent Plasticity (STDP), a neurobiological learning rule where the strength of association between two events depends on their relative timing. The study adapts this principle to structured EHR data to create sparse, directional transition features.

• Feature Construction:

  • Event Representation: Structured EHR data (diagnoses, abnormal lab values, procedures) are mapped into timestamped clinical event families. Lab values are grouped into interpretable physiological domains (e.g., renal, hepatobiliary, inflammatory).
  • STDP Transition Scoring: For any ordered pair of event types $(i, j)$ occurring at times $t_i$ and $t_j$, a transition feature is generated if $i$ precedes $j$ within a prespecified lag window $L$.
  • Weighting: The contribution of a qualifying pair is weighted by an exponential decay function, $w_{ij} = \exp(-(t_j - t_i)/\tau)$, where $\tau$ controls the decay rate. This ensures that temporally proximal transitions contribute more strongly than distant co-occurrences. Directionality is preserved ($i \to j$ is distinct from $j \to i$).
  • Aggregation: Transition scores are aggregated for each patient-snapshot to form a feature vector, which is concatenated with static burden features (demographics, cumulative counts, baseline labs).

• Study Design & Datasets:
  The framework was evaluated in two distinct settings using patient-level cross-validation and rolling prediction horizons:

  1. Acute Kidney Injury (AKI): A dense, short-term setting using 17,351 ICU stays from MIMIC-IV. The task was to predict incident AKI within 48 or 72 hours based on data from the first 24 hours of admission.
  2. Pancreatic Ductal Adenocarcinoma (PDAC) Recurrence: A sparse, long-term setting using 713 patients from Columbia University Irving Medical Center (CUMC). The task was to predict early postoperative recurrence (within 6 months) using preoperative EHR history, with rolling snapshots from Day -90 to Day 0 relative to surgery.

• Models: The study compared three model families: (1) Static burden baselines, (2) Single-STDP transition models, and (3) Calibrated STDP ensembles (combining static and transition features). Algorithms included logistic regression, elastic net, XGBoost, and LightGBM.

Key Results

• AKI Prediction (Dense Setting): The calibrated STDP ensemble significantly outperformed static burden models.
  • At the 48-hour horizon, the STDP ensemble achieved an AUROC of 0.868 compared to 0.827 for the static baseline.
  • At the 72-hour horizon, the STDP ensemble achieved an AUROC of 0.838 versus 0.800 for the static baseline.
  • Improvements were also observed in AUPRC, Brier score, and sensitivity at high specificity (90%). Feature attribution highlighted clinically plausible pathways, such as cardiovascular and renal transitions preceding AKI.
• PDAC Recurrence Prediction (Sparse Setting): The STDP ensemble showed modest but consistent improvements over static baselines.
  • At Day -30 (primary decision horizon), the STDP ensemble achieved an AUROC of 0.611 versus 0.587 for the static baseline.
  • At Day 0, the AUROC was 0.681 versus 0.657.
  • The gains were smaller than in the AKI task, reflecting the sparsity of oncology surveillance data and the longer biological timescale of recurrence.
• Robustness: Analyses confirmed that performance gains were not solely driven by observation intensity (e.g., number of lab tests or visit frequency). STDP models maintained superiority even after adjusting for record length and encounter density.

Significance and Claims
The paper claims that STDP-inspired transition features offer a practical and interpretable method for incorporating temporal ordering into EHR-based risk prediction without the "black box" nature of deep learning sequence models.

• Interpretability: Unlike latent embeddings, the STDP features explicitly encode event identity, directionality, and approximate timing, allowing clinicians to identify specific temporal trajectories (e.g., "liver abnormalities followed by nutritional decline") that drive risk.
• Versatility: The framework is effective across both acute, high-density settings (ICU) and sparse, longitudinal settings (oncology surveillance), though the magnitude of improvement correlates with event density and temporal proximity.
• Clinical Utility: The approach supports rolling prediction horizons, enabling continuous risk stratification. In the PDAC context, the authors emphasize that the model is intended to support tumor-board discussions and surveillance intensity rather than standalone surgical decision-making, particularly given the need for high specificity to avoid false-positive high-risk classifications.

The authors conclude that while the method is not a mechanistic model of biological plasticity, it serves as a robust computational analogy that preserves temporal structure while retaining feature-level interpretability, bridging the gap between static burden models and complex temporal deep learning.