Create a healthier world with AI

December 2021

We explore the feasibility of detecting embedded dictations in recordings of doctor-patient conversations in this research. A viable approach to practically segment audio recordings to highlight dictations is presented.

November 2021

Medically minor but clinically important findings associated with physical child abuse may be identified by frontline clinicians yet the association of these injuries with child abuse is often not recognized. An accurate natural language processing (NLP) algorithm to identify high-risk injuries in electronic health record notes could improve detection and awareness of abuse.

July 2022

We show that pre-training of transformer models with in-domain medical conversations leads to performance gains for the task of generating the history of present illness (HPI) section of clinical notes from doctor-patient conversation transcripts. Additionally, we find that long-sequence transformers are more effective at extracting relevant information than conventional models like BART.

September 2022 

This paper explores two techniques to leverage acoustic modeling data for large vocabulary speech recognition to improve a purpose built wake word detector: transfer learning and knowledge distillation. See how these techniques interact with time synchronous training targets to improve detection latency.

September 2022 

This paper reviews a structured information retrieval and topic oriented clinical notes generation from doctor-patient conversations.

December 2022 

This work provides a systematic comparison of different decomposition methods for NLI-based factuality scoring of abstractive summaries. We highlight shortcomings with existing factuality scoring methods and describe techniques to improve performance.

November 2022

We first revisit the calibration properties of regular focal loss to highlight the downside of using a single fixed γ for all the training samples, and find that there's no single γ that can achieve the best calibration and propose a calibration-aware adaptive strategy called AdaFocal, that updates the γ at every step based on γ from the previous time step and the magnitude of the model’s under/over-confidence. AdaFocal performs consistently better across different dataset-model pairs producing inherently calibrated models that benefit further from post-hoc calibration techniques.

July 2023

We introduce MDACE, the first publicly available dataset for evidence/rationale extraction on an extreme multi-label classification task over long medical documents. The dataset is built on a subset of the MIMIC-III clinical records and was annotated by professional medical coders.  

August 2023 

We explore using synthetic data in pre-training sequence-to-sequence (seq-to-seq) transformer models to generate clinical notes from Doctor-Patient-Conversations (DoPaCos). Using a generative language model fine-tuned on authentic conversations, a synthetic DoPaCo dataset was created and used with a corpus of clinical notes to pre-train a Longformer-Encoder-Decoder (LED) model.