Clinical ML at DASION

DASION is a clinical AI startup founded with Prof. Weiqing Gu (Harvey Mudd) and supported by NSF SBIR Phase I and II awards. I joined in 2021 as an intern and have been a Machine Learning Engineer since June 2025.

Depression detection. A multimodal multi-instance model over clinical interview audio and transcripts: frozen DepRoBERTa and emotion-finetuned wav2vec2 encoders, gated-attention pooling over interview chunks, and an occlusion diagnostic that reports how much each modality contributed. F1 0.87 on DAIC-WOZ. Written up as a solo-author paper now under review.

Imaging and documentation. ResNet50 transfer-learning classifiers with SHAP explanations for lung and cardiac imaging (93% diagnostic accuracy on small clinical datasets), and BERT fine-tuned for clinical documentation classification that cut manual processing time by 60%.

Deployment. Models are served as AWS inference endpoints (SageMaker, Lambda, S3) in Docker to five-plus partner institutions.