Patient foundation model for risk stratification in low-risk overweight patients

TL;DR

PatientTPP, a neural temporal point process model trained on over 500,000 real-world clinical trajectories to learn patient representations from sequences of diagnoses, labs, and medications, offers an interpretable, general-purpose foundation for patient risk modeling with direct applications to obesity-related care and cost targeting.

Abstract

Accurate risk stratification in patients with overweight or obesity is critical for guiding preventive care and allocating high-cost therapies such as GLP-1 receptor agonists. We present PatientTPP, a neural temporal point process (TPP) model trained on over 500,000 real-world clinical trajectories to learn patient representations from sequences of diagnoses, labs, and medications. We extend existing TPP modeling approaches to include static and numeric features and incorporate clinical knowledge for event encoding. PatientTPP representations support downstream prediction tasks, including classification of obesity-associated outcomes in low-risk individuals, even for events not explicitly modeled during training. In health economic evaluation, PatientTPP outperformed body mass index in stratifying patients by future cardiovascular-related healthcare costs, identifying higher-risk patients more efficiently. By modeling both the type and timing of clinical events, PatientTPP offers an interpretable, general-purpose foundation for patient risk modeling with direct applications to obesity-related care and cost targeting.

Figures (12)

• Figure 1: A. A temporal point process is an event sequence defined over some interval $[0, T)$. Events arrive at real-valued timestamps on this interval. Events may have classes associated with them (also known as "marks") but not magnitudes (at least not out-of-the-box). Events have no duration. The process is presumed to be autoregressive, that is, past performance is some indication of future behavior. B. A TPP model learns a conditional intensity function, which represents the joint probability distribution over all events. Interdependencies between all event types are potentially captured. There are no restrictions on data spacing or sparsity.
• Figure 2: A. PatientTPP extensions to the AttNHP encoder architecture. Input sequence events and timestamps are embedded separately, then concatenated and fed to a multi-head self-attentive transformer. Invariant features are added in the main context-building loop are are nonzero for the first item in the sequence. Numeric features pass through their own TPP submodel (yellow) and are concatenated with the indicative feature results and the clinical embeddings. The unified stream is transformed by two fully connected layers and a Softplus activation, yielding intensity scores. The output has dimensions $[batch size] {\times} [maximum sequence length] {\times} [hidden layer size]$. B. Patient characterization and TPP inference. Hidden layers of the trained TPP model encode the patient's context and functionally serve as an embedding space for EHR, simplifying cohort identification and risk assessment. We pool the contexts for six predicted conditions to arrive at a vector representation of a patient.
• Figure 3: Outcome prediction from PatientTPP representations. For each outcome, a binary logistic regression was trained on PatientTPP representations for the presence or absence of the outcome in the period after embedding. Outcomes are organized into cardiovascular (left), metabolic (center), or other (right). Average area under the curve is reported for 6-fold hold-out test.
• Figure 4: Feature importance for outcome prediction. Each bar represents a regression coefficient associated with a PatientTPP embedding dimension corresponding to a specific clinical event type (y-axis). These dimensions reflect the learned intensity of events used during TPP model training. Only features with nonzero coefficients for the given outcome are shown. Bar color indicates coefficient sign (blue = positive association, red = negative), and shading reflects consistency: darker bars appeared in more of the testing folds (max = 6).
• Figure 5: Comparison of BMI and PatientTPP-predicted CV costs for patient stratification. A. Future CV-associated costs for all low-risk overweight/obese cohorts were ranked by BMI and CV-cost predicted from PatientTPP representations. B. Cumulative cost capture curve for the same patients when ranked by BMI and predicted CV-cost.