Clinical DriftOps Platform ? Honest AI for Healthcare

---

Live Dashboard: https://sobcza11.github.io/clinical-driftops-platform/

Governance-first MLOps framework ensuring transparent, auditable, and policy-as-code validation for clinical AI models.

PMI-CPMAI Aligned ? MLOps ? Trustworthy AI ? Explainable Healthcare Models

---

Overview

Clinical DriftOps Platform is an end-to-end MLOps framework for trustworthy, explainable, and continuously validated clinical AI. It monitors data drift, fairness, and model explainability, automatically enforcing governance policies via CI/CD. Built for clinical research using the MIMIC-IV dataset, aligned to PMI-CPMAI Phases I-VI.

---

CPMAI / CRISP-DM Alignment

Phase	Description	Artifacts
I. Business Understanding	Define risk of model drift & bias in clinical ML.	README.md, docs/overview.md
II. Data Understanding	Profile MIMIC-IV labs, vitals, outcomes.	src/data_prep.py, EDA notebooks
III. Data Preparation	Scale & clean baseline vs. current datasets.	data/data_prepared_*.csv, reports/data_prep_meta.json
IV. Modeling & Drift Detection	PSI / KS tests + SHAP explainability.	src/monitors/drift_detector.py, src/explain/shap_summary.py
V. Evaluation & Governance Gate	Fairness, performance metrics, policy enforcement.	src/eval/*, src/ops/policy_gate.py, reports/*
VI. Operationalization	Live monitoring, MLflow registry, FHIR integration.	src/ops/*, reports/index.html

---

Key Features

• Automated Performance Audits • AUROC, accuracy@0.5, KS via performance_metrics.py.
• Data Drift Monitoring • PSI / KS tests; thresholds in policy.yaml.
• Fairness Audits • per-group positive-rate & parity gap (fairness_audit.py).
• Explainability via SHAP • top features & artifact presence checks.
• Policy Gate Enforcement • single source of truth for model acceptance criteria.
• CI/CD • GitHub Actions runs the pipeline and publishes a live dashboard.
• MLflow Logging • metrics & artifact tracking (Phase VI expansion).

---

Policy Schema (policy.yaml)

drift:
  psi_fail: 0.20
  ks_fail: 0.20

performance:
  min_auroc: 0.80
  # optional: min_auprc: 0.50
  # optional: max_log_loss: 0.70

fairness:
  parity_gap_fail: 0.05

explainability:
  require_shap_artifact: true
  top_features_min: 10