Healthcare Applications of Time Series Foundation Models

Healthcare generates an extraordinary volume of time series data. Continuous ECG waveforms, minute-by-minute vital signs (heart rate, SpO2, blood pressure), continuous glucose monitor readings, EEG signals, and operational metrics like bed occupancy and patient flow all produce temporal streams that clinicians must interpret under time pressure. Machine learning has always promised to help, but adoption has been held back by a stubborn reality: labeled medical data is extremely scarce and expensive to produce. Time series foundation models offer a path through this bottleneck.

Why Healthcare Needs Zero-Shot Transfer

Traditional supervised learning requires large volumes of annotated training data. In healthcare, producing annotations demands expert clinician time — labeling an arrhythmia requires a cardiologist, confirming sepsis onset requires retrospective chart review by a critical care physician. These costs make it prohibitively expensive to build per-task supervised models for every clinical use case.

Zero-shot forecasting sidesteps this bottleneck. A TSFM pretrained on diverse time series data learns general temporal patterns — periodicity, trend, autocorrelation, regime changes — that transfer to clinical time series without task-specific training. A model that understands cyclical patterns from industrial sensors can recognize the quasi-periodic rhythm of a heart rate signal. A model trained to detect distribution shifts in server metrics can identify the subtle vital sign drift preceding patient deterioration. This makes TSFMs particularly relevant for smaller hospitals that lack the data infrastructure of major academic medical centers. For a deeper comparison, see our guide on fine-tuning vs. zero-shot.

Clinical Use Cases

Early warning for patient deterioration. TSFMs can forecast expected vital sign trajectories and flag deviations indicating physiological instability. This applies the forecast-residual anomaly detection framework directly: if observed heart rate falls outside the model's predicted distribution, something clinically meaningful may be happening.

Sepsis onset detection. Early sepsis presents subtly — small shifts in temperature, heart rate, and white blood cell count that are individually unremarkable but collectively significant. TSFMs can detect coordinated distributional shifts across vital sign streams that precede overt sepsis criteria, buying clinicians critical hours.

Arrhythmia classification from ECG. MOMENT, the multi-task TSFM from CMU, is especially relevant here. MOMENT supports classification, anomaly detection, and imputation within a single architecture — meaning one model can classify arrhythmia types from waveform segments, flag anomalous rhythms in continuous monitoring, and impute missing data from electrode disconnection events.

Continuous glucose monitoring prediction. Predicting glucose levels 30 to 60 minutes ahead enables proactive insulin dosing rather than reactive correction. TSFMs produce not just point forecasts but prediction intervals that quantify uncertainty around predicted trajectories — critical for dosing decisions where both hyperglycemia and hypoglycemia carry risk.

Hospital resource planning. Bed occupancy, ED arrivals, and surgical scheduling produce time series that drive operational decisions. Forecasting patient flow 24 to 72 hours ahead optimizes staffing and reduces costly diversions. This parallels the energy demand forecasting problem and benefits from the same TSFM approaches.

Explainability and Regulatory Challenges

Healthcare operates under regulatory constraints most industries do not face. The FDA's Software as a Medical Device framework requires evidence of safety and efficacy. HIPAA imposes strict data handling requirements. And clinicians rightly demand explainability — a deterioration alert must provide enough context for a nurse or physician to understand and act on it.

Prediction intervals help close this gap. Rather than a binary alert, a probabilistic TSFM can show that a patient's heart rate is at the 98th percentile of its expected distribution, displaying the predicted envelope alongside the actual trajectory. This is more interpretable than a black-box score because it is grounded in the patient's own temporal context.

Calibrated uncertainty is non-negotiable. A prediction interval claiming 95% coverage that achieves only 80% will lead to missed deterioration events. Rigorous calibration testing against held-out clinical data — from sources like the MIMIC-III or eICU datasets on PhysioNet — is essential before any clinical deployment.

Practical Adoption Strategy

Given the stakes and regulatory complexity, we recommend a staged approach.

Start with operational forecasting. Hospital resource planning carries lower clinical risk than patient-facing decision support. These are forecasting problems with clear accuracy metrics that produce immediate ROI through better resource utilization.

Validate on public clinical datasets. Before moving to clinical decision support, benchmark performance on established PhysioNet datasets. MIMIC-III and eICU provide real ICU vital sign data with outcome labels, enabling rigorous evaluation without the complexity of a prospective clinical study.

Layer into clinical workflows as decision support, not automation. The near-term role for TSFMs in clinical settings is to surface information for clinician review. An early warning alert that flags a patient for nursing assessment is decision support. Automated medication adjustment is a far higher regulatory bar.

Healthcare is where the core strengths of TSFMs — zero-shot generalization, probabilistic outputs, multi-task flexibility — align directly with the domain's hardest challenges. Explore the models suited for these applications in our model catalog, or review the API documentation to begin prototyping.