TriAID – Chest X-Ray Triage System
AI-powered clinical decision support system for follow-up imaging recommendations.
Overview
TriAID is a multimodal clinical decision-support system designed to assist radiologists in identifying appropriate follow-up imaging recommendations from chest X-ray studies.
The system combines:
- computer vision outputs from MedRAX (pathology probabilities)
- structured patient data (age, sex, clinical indication)
- natural language processing of radiology reports
These signals are fused into a multi-label classification model that predicts whether a patient should receive:
- follow-up X-ray
- CT scan
- or no further imaging
The model is trained on ~26,000 real-world radiology cases and achieves strong performance with an F1 score of ~80%, demonstrating its ability to provide reliable and interpretable recommendations.
Beyond prediction, TriAID integrates into a full-stack application with:
- a backend API for inference
- a database layer for storing recommendations
- and a conversational assistant that explains model outputs
This allows the system to function not just as a model, but as a deployable clinical support tool.
Key Features
- Multimodal ML (vision + structured + text)
- Interpretable OvR classification framework
- Threshold-tuned predictions for clinical safety
- End-to-end pipeline (data → model → app)
- Explainable AI via LLM assistant
- Designed for real clinical workflow integration
Demo
▶ Click to Play Demo
Links
Reflection
This project highlighted the importance of designing machine learning systems that go beyond predictive accuracy and align with real-world constraints.
A key learning was balancing recall and precision in a clinical setting, where missing a necessary follow-up can be significantly more harmful than over-recommending one. This required careful threshold tuning and evaluation beyond standard metrics.
Additionally, working with multimodal data reinforced the value of combining heterogeneous signals. While individual components (e.g., text or imaging) provided useful information, their integration produced more robust and context-aware predictions.
Finally, building the application layer emphasized that impactful ML systems must be deployable, interpretable, and user-focused, not just accurate.
Limitations
- Strong class imbalance (majority = no follow-up) affects minority class performance
- Variability in radiology report language introduces noise in text features
- Model performance on X-ray and CT recommendations is lower than “no follow-up” cases
- Dependent on MedRAX preprocessing constraints (e.g., image formatting)
- Not clinically validated intended as a support tool, not a replacement