DT4PCP v2.0: A Digital Twin Framework for Personalized Care Planning

DT4PCP v2.0 turns the original DT4PCP framework into a deployable clinical decision-support tool for patients with Type 2 diabetes. Rather than a conceptual pipeline, it is a guided, four-tab application built around a single patient encounter.

New Visit loads an encounter and opens a session for it. Baseline predicts the probability of an emergency-department visit. It covers three windows: 0 to 30, 30 to 60, and 60 to 90 days. Paired XGBoost and Random Forest models compute each prediction. A per-feature SHAP breakdown shows what drives the risk up or down.

Simulation lets the clinician pick which factors are actually adjustable. These can include vitals, measures, and resolvable diagnoses. The clinician sets a range for each adjustable factor. The app then runs a D-optimal design of experiments. It fits a local response surface and optimizes within those ranges. The result is re-confirmed through the real model. This surfaces the scenarios most likely to lower risk. Recommendations turns the top-ranked scenarios into a clinical narrative. An LLM, Claude, generates this narrative. A field lets the clinician record their own final decision. The underlying models are trained on a Type 2 diabetes patient population. They incorporate social determinants of health, like income and insurance coverage. Distance to care is included as well. Clinical history and utilization patterns are also part of the model.

Every encounter a clinician opens becomes a persistent session. Edits, ranges, results, and decisions save automatically. They are there again when the clinician returns. Sign-in is optional, and it keeps each clinician's worklist separate. As with v1.0, DT4PCP v2.0 is a research tool. It supports clinical decisions but does not replace them. Its predictions and simulated strategies are model output, not medical advice. A qualified clinician should review them before they inform care.