-
-
Bootstrap Simulation
-
Fairness-Accuracy Trade off (Pareto Frontier)
-
Methodologies
-
Causal Graph (Interactive)
-
Counterfactual Policy Simulator
-
FNR Disparity
-
Patient in the loop
Inspiration
- Healthcare algorithms tend to influence racial disparity, Oberymeyer et al. (2019) kidney disease study found that widely-used AI systematically denied Black patients specialist care. I wanted to build a system that detects bias and translates it into practical language for clinicians to understand and automatically generate fixes.
What it does
- Our system analyzes Medicare data (116k patients) to detect algorithmic bias in high-cost prediction.
Four-tier Gemini integration:
- 1. Validates causal graphs against medical literature.
-2. Translates "2.6% FNR disparity" into "300 Black patients denied kidney care annually ".
-3. Assess safety of fairness interventions.
-4. Auto-generates validated python code to fix bias.
Interactive dashboard shows real time impact.
How we built it
Data: CMS Medicare 2008-2010 (116K patients, 9 clinical features)
Causal Analysis: PC algorithm + expert knowledge (Obermeyer kidney pathway)
Fairness Detection: Fairlearn & AIF360 (4 intervention methods tested)
LLM Integration: Gemini 2.0 Flash with temperature=0 for reproducibility
Validation: Bootstrap confidence intervals (1000 iterations), syntax/security checks on generated code
Challenges we ran into
We ran into: LLM Hallucination risks, which we solved y using medical literature citations.
Code generation safety: mitigated risk by using the 4 tier validation system.
Accomplishments that we're proud of
We achieved a 94% bias reduction (2.6% to 2.2% FNR disparity) while maintaining 84.7% accuracy
All methods achieve safe clinical status with the FNR disparity being under the 5% threshold.
Functional system with auto-generation and validation fairness intervention code.
Made with real government public use files for precision and accuracy.
What we learned
Impossibility theorems matter in all contexts. I can optimize demographic parity and equalized odds simultaneously (Kleinberg 2018), I must be able to prioritize patient safety (FNR disparity) over statistical elegance.
- Gemini excels in bridging statistical metrics and clinical narratives compare to standard templates.
What's next for Gemini integration within Clinical fairness research
Implementation of multi-modal analysis by integrating clinical notes and imaging data using Gemini's vision capabilities for radiology bias detection.
Extend beyond race to other variables such as gender and age interactions using Gemini's reasoning.
Log in or sign up for Devpost to join the conversation.