Krithi Meda

Drug Sensitivity Predictor

● Developed a multiclass drug response classifier (Sensitive/Intermediate/Resistant) on the GDSC dataset (100,000+ samples) using a classification-based approach as an alternative to the conventional regression methods. ● Engineered a balanced target variable from the continuous IC50 values using quantile-based binning, performed feature selection to eliminate data leakage and redundant genomic descriptors, and applied one-hot encoding across a high dimensional sparse feature space of 100,000+ samples. ● Compared Logistic regression, Random Forest, and XGBoost through rigorous evaluation using accuracy, F1 score, Precision-Recall curves, ROC-AUC curves, and stratified 5-fold cross-validation, selecting Logistic Regression based on F1 score and linear feature-target relationships inherent to genomic data. ● Deployed an end-to-end web application using Flask supporting single sample and batch CSV prediction of drug outcome responses.

Multimodal Disease Prediction From Chest X - rays and Clinical Notes

● Building a late-fusion multimodal pipeline in PyTorch that combines ResNet18 image embeddings (512-dim) and DistilBERT report embeddings (768-dim) through a concatenation-based MLP head, using the Indiana University (OpenI) chest X-ray dataset. ● Fine-tuned DistilBERT on radiology reports (98% accuracy) and removed train/test leakage from duplicate report text, then confirmed through manual error analysis that the model was learning real report patterns rather than memorizing. ● Fixed a severe class imbalance (80/20) in the X-ray classifier with weighted oversampling, raising Normal-class recall from 18% to 68%. ● Implemented Grad-CAM with PyTorch hooks to visualize the image regions driving the fused model’s predictions.