DOWNLOAD: Full Paper or Presentation

Rethinking Clinical AI: Patient Similarity for SSI Prediction

Click Here for TLDR; 5 Minute Audio Overview

The Critical Challenge of Surgical Site Infections

Surgical Site Infections (SSIs) are a formidable threat in healthcare, imposing severe consequences on patient well-being and creating substantial financial and operational burdens for hospitals.

$3.5-10B
Estimated Annual Cost in the U.S.

SSI's Share of Hospital Infections

SSIs account for a significant portion of all healthcare-associated infections (HAIs), making their prevention a top priority for patient safety.

2-11x
Higher Risk of Post-Operative Death

The Gap in Traditional SSI Detection

Traditional methods for detecting and predicting SSIs are often too slow. Data from labs or manual chart reviews arrive long after an infection has begun, delaying critical interventions and limiting the effectiveness of care.

The Lag Between Infection and Action

Surgery

Patient undergoes procedure

Infection Onset

Subtle signs may appear

Data Lag

Waiting for lab results, chart reviews

Late Detection

Infection is finally confirmed

This delay creates a critical "intervention gap" where proactive care is impossible. An AI model that predicts risk in real-time can bridge this gap.

The "Data-First" Patient Similarity Framework

Instead of building a more complex model, our approach engineers a better dataset. By finding non-SSI patients who are clinically similar to SSI patients, we create a balanced and highly relevant training cohort using only real-time EMR data.

How It Works: From Raw Data to Refined Cohort

🏥

1. Raw Population

Start with all 151,733 surgical cases (severely imbalanced).

🧬

2. Apply Similarity

Use Jaccard & Cosine metrics on Dx/PCS codes to compare non-SSI to SSI cases.

🎯

3. Filter & Select

Keep only non-SSI cases with a similarity score > 0.70.

📊

4. Balanced Cohort

Result: a refined set of 15,339 clinically relevant cases for model training.

Achieving High-Performance, Real-Time Prediction

The model trained on our similarity-filtered cohort demonstrates powerful predictive performance, outperforming numerous other standard classifiers—all without using slow, time-lagged data like lab results, imaging, or clinician notes.

Model Performance Comparison (AUC)

The CatBoost Classifier emerged as the top-performing model, demonstrating superior ability to distinguish between SSI and non-SSI cases.

CatBoost Classifier: A Balanced Profile

With a high AUC of 0.9052 and a strong balance between Precision and Recall, the model is both accurate and reliable for clinical use.

Detailed Model Results

Model Accuracy AUC Recall Precision F1-Score
CatBoost Classifier.8463.9052.7317.8044.7663
Light Gradient Boosting Machine.8416.9029.7349.7906.7616
Extreme Gradient Boosting.8407.9001.7301.7913.7593
Random Forest Classifier.8198.8823.7309.7422.7363

Comparing AUC: This Study vs. Prior Research

Our model's performance (AUC 0.9052) aligns with the top results from recent literature, demonstrating competitive predictive power with fewer, more actionable features.

Strategic Impact for Clinical AI

This "data-first" approach provides a blueprint for developing more effective, practical, and trustworthy AI systems in healthcare, moving beyond theoretical performance to real-world impact.

⏱️

Real-Time Feasibility

Uses only contemporaneously available EMR data, enabling early, proactive interventions that can save lives and reduce costs.

💡

Interpretable & Trusted

The similarity logic is simple and aligns with clinical intuition, making the model easier to understand, trust, and adopt.

🚀

Scalable & Generalizable

The focus on common structured data makes the framework applicable across diverse hospital settings and systems.