Foundation Model Forecasting
1. Nightly RMSSD and HR - Chronos-2 Probabilistic Forecast
55 nights context, 14 nights forecast. Shaded bands: 90% (light) and 50% (dark) prediction intervals. Cross = anomaly outside 90% PI.
2. Continuous HR - Hourly Resolution Forecast
21K+ HR readings downsampled to hourly means. Sliding window detection identifies anomalous segments.
3. Ensemble Consensus - Foundation Model + ARIMA Baseline
Two independent models score each date. HIGH = both flag anomaly, MEDIUM = one, NORMAL = none.
4. Feb 9 Retrospective Validation
Model trained on all data through February 8, forecast for February 9-15. Tests whether the foundation model would have flagged the acute event prospectively.
| Method | Series | Detected? | Residual |
|---|---|---|---|
| Chronos Retrospective | RMSSD | YES | -5.28 |
| Chronos Retrospective | HR | YES | 13.04 |
Prospective March ensemble HIGH-confidence anomalies:
None detected
Feb 9 in March ensemble consensus:
No
5. Pre vs Post Ruxolitinib Regime Analysis
Model trained on pre-ruxolitinib period and forecasts into treatment period. Narrower prediction intervals indicate stabilization, systematic shift indicates treatment effect.
Detailed Metrics
| Section | Metric | Value |
|---|---|---|
| chronos_hourly_hr | ||
| total_hours | 5868 | |
| context_hours | 2048 | |
| forecast_hours | 48 | |
| mae | 8.330 | |
| residual_std | 8.810 | |
| coverage_90pi | 58.300 | |
| n_anomalous_hours_2sd | 3 | |
| sliding_window_anomalies | 168 | |
| feb9_hourly_anomalies | 9 | |
| inference_time_s | 0.080 | |
| chronos_nightly_hr | ||
| context_nights | 55 | |
| forecast_nights | 14 | |
| context_start | 2026-01-08 | |
| context_end | 2026-03-03 | |
| forecast_start | 2026-03-04 | |
| forecast_end | 2026-03-17 | |
| mae | 4.239 | |
| rmse | 5.389 | |
| mape | 4.750 | |
| coverage_90pi | 85.700 | |
| coverage_50pi | 50.000 | |
| pi_width_90_mean | 17.310 | |
| pi_width_50_mean | 8.740 | |
| n_anomalies_outside_90pi | 2 | |
| anomaly_dates | 2026-03-05, 2026-03-15 | |
| inference_time_s | 0.050 | |
| chronos_nightly_rmssd | ||
| context_nights | 55 | |
| forecast_nights | 14 | |
| context_start | 2026-01-08 | |
| context_end | 2026-03-03 | |
| forecast_start | 2026-03-04 | |
| forecast_end | 2026-03-17 | |
| mae | 1.387 | |
| rmse | 1.758 | |
| mape | 11.150 | |
| coverage_90pi | 100.000 | |
| coverage_50pi | 78.600 | |
| pi_width_90_mean | 8.770 | |
| pi_width_50_mean | 4.180 | |
| n_anomalies_outside_90pi | 0 | |
| anomaly_dates | ||
| inference_time_s | 0.080 | |
| data_range | ||
| start | 2026-01-08 | |
| end | 2026-10-06 | |
| n_nights | 274 | |
| ensemble_consensus | ||
| high_confidence_anomaly_dates | ||
| medium_confidence_count | 3 | |
| normal_count | 25 | |
| feb9_detected | No | |
| feb9_retro_hr | ||
| context_nights | 32 | |
| forecast_nights | 7 | |
| forecast_dates | 2026-02-09, 2026-02-10, 2026-02-11, 2026-02-12, 2026-02-13, 2026-02-14, 2026-02-15 | |
| feb9_detected | YES | |
| feb9_residual | 13.040 | |
| feb9_direction | above | |
| feb9_actual | 109.760 | |
| feb9_median_forecast | 96.720 | |
| n_anomalies_in_window | 6 | |
| detection_rate_pct | 85.700 | |
| inference_time_s | 0.050 | |
| feb9_retro_rmssd | ||
| context_nights | 32 | |
| forecast_nights | 7 | |
| forecast_dates | 2026-02-09, 2026-02-10, 2026-02-11, 2026-02-12, 2026-02-13, 2026-02-14, 2026-02-15 | |
| feb9_detected | YES | |
| feb9_residual | -5.280 | |
| feb9_direction | below | |
| feb9_actual | 4.990 | |
| feb9_median_forecast | 10.270 | |
| n_anomalies_in_window | 2 | |
| detection_rate_pct | 28.600 | |
| inference_time_s | 0.080 | |
| ruxolitinib_hr | ||
| pre_period_nights | 67 | |
| post_period_nights | 207 | |
| pre_pi_width_90 | 17.840 | |
| post_pi_width_90 | 19.040 | |
| uncertainty_change_pct | 6.700 | |
| inference_time_s | 0.310 | |
| ruxolitinib_rmssd | ||
| pre_period_nights | 67 | |
| post_period_nights | 205 | |
| pre_pi_width_90 | 5.870 | |
| post_pi_width_90 | 10.650 | |
| uncertainty_change_pct | 81.400 | |
| inference_time_s | 0.280 | |
| statistical_hr | ||
| model | ARIMA(0, 1, 2) | |
| mae | 4.270 | |
| rmse | 5.493 | |
| coverage_90ci | 100.000 | |
| n_anomalies | 0 | |
| anomaly_dates | ||
| statistical_rmssd | ||
| model | ARIMA(0, 1, 1) | |
| mae | 1.448 | |
| rmse | 1.843 | |
| coverage_90ci | 92.900 | |
| n_anomalies | 1 | |
| anomaly_dates | 2026-03-09 | |