BP Prediction Visualizer

This page is a working sketch of an idea: if we have a few days of recent blood-pressure readings, can a small, transparent model say something sensible about the next forty-eight hours? The answer here is “sometimes”—and that’s deliberate. You’ll see the familiar daily rhythm: lower values while sleeping, a rise after waking, and small wobbles across the day. The forecast traces that rhythm forward and adds a shaded band to show uncertainty. When the band is wide, we’re telling you the model is unsure.

Two simple baselines sit under the hood. The first is an ARX(1) model that remembers yesterday’s level and adjusts for the time of day—think “short memory plus a day/night heartbeat.” The second is a tiny gradient-boosted model that learns a few bends and kinks in the curve from recent history. Neither is meant to be a final word; they’re here because their assumptions are easy to see, easy to question, and easy to improve.

Real life doesn’t stay tidy, so the toggles let you play director: miss a dose, have a salty meal, or go through a stressful patch and watch the short-term forecast bump upward. If you take a new reading, use “recalibrate” and the models will fold it in; you’ll often see the uncertainty band tighten afterwards. The small tiles on the left report a recent error metric (MAE) and a plain-English confidence badge. High error or sparse data will push confidence down, as it should.

Important: this is a research demonstrator. It is not a medical device, and it shouldn’t be used to diagnose, treat, or change medications. Any real deployment would need calibration to approved cuffs, larger and more diverse datasets, careful handling of medications and illness, and clinical oversight.

A practical system would weave in more context—wearable activity and sleep, medication timing, travel and shift-work patterns—and personalise the model to you over time. It would favour uncertainty-aware forecasts, show clear “no forecast” states when the signal is poor, and surface gentle, actionable prompts like “take a reading this evening.” Most importantly, it would continuously evaluate fairness and accuracy across age, sex, ethnicity, and device types, with privacy-preserving data practices by default.