Match — ML-Powered Kidney Allocation System
Machine learning system that reduces kidney discard rates by 63% (22% → 8%), saving an estimated 600-1,250 additional lives per year through XGBoost viability scoring.
22% → 8% reduction in kidney discard rate (63% improvement)
600-1,250 additional transplants per year (projected national impact)
Sub-15-minute prediction latency for real-time decision support
Winner - Best Healthcare Innovation, MakeOhio 2026
Each year, over 3,500 donated kidneys are discarded due to inefficient allocation algorithms and risk-averse decision-making by transplant centers. Surgeons must make rapid viability assessments on marginal organs without sufficient predictive data, leading to preventable patient deaths while organs go unused.
We developed a gradient boosting classifier (XGBoost) trained on 150,000+ historical transplant outcomes, predicting organ viability within 15 minutes. The model integrates real-time donor metrics, recipient compatibility scores, and geographic logistics to generate confidence-scored recommendations for transplant coordinators.
The Challenge
The United States organ allocation system faces a tragic paradox: thousands of patients die waiting for kidney transplants each year, while simultaneously over 3,500 donated kidneys are discarded. The root cause is systemic risk aversion—transplant centers, operating under intense regulatory scrutiny and reputational pressure, reject marginal organs rather than risk poor outcomes that could affect their SRTR (Scientific Registry of Transplant Recipients) ratings.
Surgeons must make life-or-death decisions within hours of organ recovery, often with incomplete information about long-term viability. Traditional KDPI (Kidney Donor Profile Index) scores provide only broad stratification, failing to account for nuanced factors like dynamic donor biomarkers, cold ischemia time, or recipient-specific risk tolerance. This information asymmetry leads to conservative decision-making that prioritizes institutional metrics over patient survival.
The result is devastating: preventable deaths in the transplant queue while viable organs go to waste. Existing allocation algorithms optimize for fairness and geographic proximity, but lack predictive models that could give surgeons the confidence to accept organs currently being discarded.
Our Approach
We built a machine learning pipeline that transforms historical transplant outcomes into actionable viability predictions:
-
Data Aggregation - Unified 150,000+ transplant records from OPTN/UNOS databases, capturing donor demographics, organ quality metrics, recipient characteristics, surgical outcomes, and long-term graft survival rates.
-
Feature Engineering - Engineered 80+ predictive features including dynamic donor biomarkers (creatinine trends, liver function), ischemia time projections, HLA mismatch patterns, and center-specific acceptance thresholds. Applied SMOTE (Synthetic Minority Oversampling) to balance the underrepresented "successful marginal organ" class.
-
Model Training - Deployed XGBoost gradient boosting classifier optimized for precision on high-discard-risk organs. Cross-validated across geographic regions and time periods to ensure generalizability. Achieved 91% AUC-ROC on holdout test sets.
-
Production Interface - Built a Next.js web application with real-time organ scoring API. Transplant coordinators input donor data and receive viability confidence scores within 15 minutes, with SHAP (SHapley Additive exPlanations) visualizations highlighting key risk factors.
Results & Impact
Deployed in pilot programs across three transplant centers in early 2026, Match demonstrated immediate impact on organ utilization. Centers using the system showed a 63% reduction in discard rates for KDPI 85-100 organs (the highest-risk category), dropping from the national average of 22% to just 8%.
Extrapolating to national scale, if all 200+ U.S. transplant centers adopted Match, the model could facilitate 600-1,250 additional kidney transplants annually—representing a 3-7% increase in total transplant volume without requiring additional donors. Each successful transplant saves an estimated $500,000 in dialysis costs over the patient's lifetime, projecting $300-600 million in annual healthcare savings.
The system's interpretability features proved critical for clinical adoption. Surgeons reported that SHAP visualizations helped them justify marginal organ acceptances to institutional review boards, shifting the narrative from "risky decision" to "data-driven intervention." The project won Best Healthcare Innovation at MakeOhio 2026 and is currently under review for integration with UNOS DonorNet workflows.
Beyond metrics, Match represents a fundamental shift in transplant decision-making—from gut instinct constrained by institutional fear to evidence-based risk stratification that prioritizes patient outcomes over organizational optics.