BiqDS: Biophysics-Informed Quantitative Diabetes Screener

This project develops a fully interpretable, biologically grounded diabetes risk screener using structured signal logic and hybrid probabilistic gating. Unlike black-box models, this system prioritizes high recall, transparency, and modular reasoning — ideal for real-world healthcare screening and research environments.

---

🧠 Key Concept

BiqDS (Biophysics-Informed Quantitative Diabetes Screener) is based on the insight that:

Early onset of diabetes is associated with a measurable reduction in projected longevity within phenotypic clusters.

Using this principle, the system builds a FinalSignal that blends:

• Cluster-level longevity estimates,
• Observed diabetes prevalence,
• Delta in life expectancy between diabetic and non-diabetic patients,
• And hybrid filters using Random Forest model confidence when needed.

---

🚀 Features

• ✅ Unsupervised phenotypic clustering (k=35) to define biological groupings
• ✅ Longevity modeling from cluster-level age averages
• ✅ Δ-longevity weights to adjust signal confidence
• ✅ FinalSignal: interpretable risk score derived from biophysical and statistical logic
• ✅ Threshold sweep and F1 optimization
• ✅ Hybrid filtering with Random Forest probability for robust screening
• ✅ High recall configuration for maximum sensitivity

---

📊 Methodology Overview

1. Data Cleaning & Clipping
   Remove impossible values and apply domain-informed outlier clipping to enhance medical validity.

2. Clustering (KMeans)
   Segment patients into biological phenotype clusters using scaled input features.

3. Longevity Modeling
   Compute average age per cluster and normalize to build a LongevityScore.

4. Diabetes Prevalence
   Measure cluster-level diabetes probability.

5. Delta-Longevity Weighting
   Estimate how much diabetes shortens life within each cluster to compute DeltaLongevityWeight.

6. FinalSignal Construction
   Blend LongevityScore and DiabetesProbability with DeltaLongevityWeight.

7. Threshold Optimization
   Sweep fixed thresholds (e.g. 0.10–0.25) to select high-recall, high-F1 configurations.

8. Hybrid Rule (Optional)
   Apply a probability veto using a Random Forest trained on correlation-ranked features.

---

📈 Example Output

Threshold	Accuracy	Precision	Recall	F1 Score
0.20	0.85	0.68	0.88	0.77
0.10 (high recall)	0.77	0.53	0.98	0.69

Confusion matrix and classification reports included in the notebook.

---

🧪 How to Run

1. Clone the repo

   git clone https://github.com/yourusername/biqds-diabetes-screener
   cd biqds-diabetes-screener
   
   Run the notebook:

BiqDS.ipynb

Dataset required: Place diabetes.csv (PIMA dataset) in the same directory.