PAD Detection + Risk Copilot

The pipeline predicts PAD from data a hospital already collects: labs, demographics, comorbidities and medications. The copilot answers why this score? with SHAP attributions and cited passages from public clinical references.

Researchers and clinicians reviewing a risk score, for research use only. This is not a medical device and has not been validated prospectively or externally.

• Data use agreement. MIMIC-IV values cannot be sent to a third party under the PhysioNet agreement, so the language model runs locally through Ollama and no API is used.
• Hardware. Both models (llama3.1:8b and nomic-embed-text) have to fit together on an 8 GB GPU.
• Noisy label. The label is a billing code, which can miss or mislabel patients, and MIMIC-IV is an ICU population, sicker than the general public.
• Missing signals. There is no ankle-brachial index, no smoking status and no imaging in these tables.

The model. A matched case-control cohort from MIMIC-IV, 13 engineered features (demographics, labs, comorbidities, medications) and six classifiers: logistic regression, random forest, SVM, MLP, XGBoost and LightGBM. The winner is chosen by grouped 5-fold cross-validation and scored once on a held-out test set.

Leakage controls, each covered by a test that was checked to fail against the earlier behaviour:

• Patient-level cohort and split, so no one appears as both a case and a control, or in both train and test.
• Comorbidities and medications come only from strictly earlier admissions. Statins are the standard PAD treatment, so counting them from the index stay would leak the diagnosis.
• Controls matched 1:1 on age bracket × gender, so the model has to learn something clinical.
• Drug names are matched exactly. A bare statin substring would also match nystatin, an antifungal.

The copilot. It scores the record, computes SHAP attributions, retrieves passages from an index of NHLBI and CDC pages plus a model card, asks a local LLM to explain citing those passages, then verifies every citation in code. Any [n] that doesn't resolve to a retrieved passage is stripped and flagged. There's a Streamlit UI on top.

The two answers without a citation are marked as ungrounded in the UI.

• A local 8B model.The data agreement rules out hosted models, so the copilot runs llama3.1:8b locally. Its answers are weaker, which is one reason citations are checked in code.
• The copilot can decline to answer.If no passage clears the relevance threshold, the LLM is never called and the copilot says so.
• The test embedder must be asked for by name.An offline hashing embedder exists for tests. It scores 0.79 / 0.54 (hit@6 / hit@1) against 1.00 / 0.79 for real embeddings, so it is never swapped in automatically.
• SHAP factors are worded as score movement.Given "statin therapy increases", the model once wrote that statins increase PAD risk. Factors are now phrased as "pushed the score up/down".
• Only prior admissions for comorbidities and medications.This gives up some signal to avoid leaking the diagnosis into the features.

• Calibrate the probabilities (Platt or isotonic). Matched controls change the base rate, so today's scores rank patients well but need calibrating before they can be read as probabilities.
• Rerank retrieved passages before generation.
• Add features from MIMIC-IV-Note clinical notes, such as smoking status and mentions of claudication.