Praveen Bansal

Invented a novel algorithm for detecting lack of compassion in Healthcare Adverse Determination Letters. Implemented DeepMoji model to represent each sentence into 64 popularly used emoticons. Leveraged these emoticon probabilities and Emoji Sentiment Ranking to generate a compassion score. This algorithm exceeded F1- score of all the existing Python/R libraries on 722 manually annotated letters. Built Generic Commercial Model for tagging fraudulent claims for 10+ different clients. For every new client, Customized model took 9 months and dedicated DS bandwidth to onboard this new. Generic model takes only 4 weeks with minimal manual intervention. The model provides USD per touch value of $ 300 which is 30% more than the existing supervised models.

Modified the YOLO - Object detection system to create the medical history of members using only 200 samples. Developed an algorithm using Transfer Learning to identify check boxes in an unknown image format without any availability of labelled data. Used Optical Character Recognition (OCR) to read the member diagnostics which can be further leveraged for creating features in waste & error models. Developed a Machine Learning model for generating waste and error leads for behavioral health claims. Currently, model is in production and is expected to give 20% incremental savings based on historical data. Developed a reason code generator, which provides explanations for model’s prediction.

Automated the standardization of unstructured Patient Plans data resulting in savings of 200 FTEs. Developed a rule-based algorithm to transform complex unstructured Census data from multiple sources to a standard format to be consumed by Newton tool to generate various reports.