Overview
Built an AI-powered healthcare insurance fraud detection platform designed for small and mid-sized health insurers, Medicaid administrators, and third-party claims processors. The consolidated case study covers two related system layers: a machine-learning decision-support dashboard and an earlier event-driven pipeline for provider-level fraud and cyber risk monitoring.
Business Problem
Healthcare claims teams need more than a fraud score. They need a workflow that can prioritize risk, route borderline cases intelligently, support investigators with explainable outputs, and avoid unsafe automation in critical financial decisions.
Datasets
- Primary dataset: Kaggle Healthcare Provider Fraud Detection Analysis
- Secondary dataset: synthetic Health Insurance Claims Data for Fraud Detection
These datasets were used to model provider and claim risk patterns, generate reusable scoring logic, and stress-test operational routing decisions.
Modeling Approach
- Built provider-level fraud features from claims and beneficiary data
- Compared logistic regression, random forest, and boosting models
- Tuned thresholds for imbalanced fraud detection
- Created reusable scoring logic for provider and claim fraud probability outputs
- Focused on explainability and business-facing interpretation rather than raw model complexity alone
Decision Automation Engine
The platform includes a decision automation layer that translates fraud probability into operational routing actions:
- Auto Approve
- Secondary Rules Check
- Human Review Queue
- Priority Investigation
- Payment Hold + Mandatory Investigator Review
This makes the system useful for day-to-day fraud operations, not just model experimentation.
Dashboard Experience
The Streamlit dashboard includes dedicated pages for:
- Overview
- Batch Scoring
- Decision Engine
- Case Review
- Model Insights
- Business Impact
- Responsible AI
- About
The product framing emphasizes investigator support, operational clarity, and executive readability.
Real-Time Risk Monitoring Layer
The project now incorporates the earlier Real-Time Healthcare Fraud and Cyber Risk Pipeline as a companion architecture rather than presenting it as a separate portfolio project.
That layer includes:
- simulated healthcare claim and cybersecurity access-log events
- Kafka / Confluent-style producers and consumers
- rule-based fraud and cyber anomaly detection
- provider-level fraud and cyber signal correlation
- SQLite alert storage
- Streamlit monitoring views
- a local demonstration mode that does not require Kafka
The streaming layer demonstrates event-driven data engineering and rapid risk alerting. The machine-learning layer focuses on fraud probability scoring, investigation workflow, business impact, explainability, and responsible decision controls. They remain distinct implementations within one healthcare risk case study rather than being presented as a fully integrated production platform.
Business Impact
The system includes business impact modeling to estimate:
- fraud prevented
- review cost
- net savings
This makes the platform easier to evaluate in terms of operational value, not only model metrics.
Responsible AI Controls
- No automatic denial language
- Human review required for critical cases
- Secondary rule checks for borderline cases
- Claims above $10,000 require human review
These controls keep the workflow aligned with responsible AI governance and reduce the risk of unsafe automated decisioning.
Tools and Technologies
- Python
- pandas
- scikit-learn
- Streamlit
- Plotly
- joblib
- Kafka-style streaming architecture
- SQLite alert storage
- provider-level feature engineering
- reusable scoring and decision-routing logic
Why It Stands Out
This project stands out because it connects applied machine learning with fraud operations workflow design, business decision support, dashboarding, explainability, and responsible AI governance.
In one system, it combines:
- applied machine learning
- fraud operations workflow design
- business decision support
- executive dashboard design
- explainability
- responsible AI governance
- streaming analytics and event-driven risk correlation
Future Hybrid Roadmap
Future versions may connect the streaming event pipeline directly to the machine-learning scoring and investigator dashboard layers to create a unified real-time healthcare risk platform.