Charité AI

During an internship at the Department of Urology at Charité – Universitätsmedizin Berlin, I developed a Python-based AI system for the text-based detection and classification of iatrogenic, surgery-related complications. The tool was built as a quality component for a planned joint study between BHT and Charité, and tested against pseudonymized patient records extracted from the hospital's clinical information system (KIS). The system trains a separate deep neural network for each of five complication types — including ureter fistula, bladder fistula, and urinoma — since each has its own distinct risk profile and class imbalance characteristics. A key challenge was the extreme rarity of some complications (< 2% of cases), which was addressed through SMOTE oversampling, per-class weight tuning, and individually calibrated decision thresholds. Selected flagged cases are subsequently reviewed by urologists for clinical validation.

Key Features

• Per-complication neural network models with individually tuned hyperparameters.
• SMOTE-based oversampling combined with class weighting to handle severe data imbalance.
• Stratified k-fold cross-validation for robust performance evaluation.
• Feature importance analysis via Random Forest to identify the strongest clinical risk factors.
• Full data preprocessing pipeline: ICD-10 code parsing, BMI calculation, median imputation, and outlier handling.
• CLI-based interface (stdin/stdout) for integration into clinical research workflows.