snaibcell

Inspiration

Stroke thrombectomy is one of the most time-critical procedures in medicine. Every minute of delay destroys neurons. But once the surgeon is inside, there's no patient-specific guidance on how long they can safely operate before the brain starts to deteriorate. Surgeons rely on experience and intuition. We wanted to change that.

What We Built

Snaibcell is a pre-operative clinical decision support tool that predicts the safe surgical window for stroke thrombectomy patients. Input a patient's clinical profile before surgery — the model tells you how long the surgeon can safely operate.

How We Built It

We built a two-stage machine learning pipeline using XGBoost. The first stage classifies bad outcomes (hemorrhagic transformation, 30-day mortality, or mRS ≥ 3). A duration sweep across each patient identifies the point where bad outcome probability exceeds 30% — this becomes the regression target. The second stage trains a regressor on that target using 56 features including 28 clinically engineered variables across brain tolerance, oxygen delivery, clot burden, and deterioration rate. A 20-minute safety buffer is applied to all predictions.

We integrated Featherless.AI (OpenBioLLM-70B) to generate a pre-operative clinical brief for the attending physician, identifying abnormal values and summarizing patient risk in clinical language. The API is deployed on Vultr and the frontend is built in React.

Challenges

• Defining a meaningful regression target without real surgical outcome data required the two-stage sweep approach
• Removing post-operative data leakage while preserving enough signal for accurate prediction
• Keeping the tool clinically useful without crossing into diagnostic or prescriptive territory

What We Learned

• Clinical feature engineering grounded in stroke medicine produces more meaningful models than raw feature lists
• Two-stage pipelines can generate useful regression targets from classification models when ground truth is unavailable
• The difference between a decision support tool and a diagnostic tool matters both ethically and clinically

Built With

• fastapi
• featherless
• python
• react
• sklearn
• tailwind
• typescript
• xgboost