Working with a top 20 global pharma company, we developed a custom NLP-based social listening & medical topic modeling system for their global patient call center providing automated company-patient communication analysis. The system leverages traditional NLP techniques, deep learning, and optimized speech recognition techniques to create a State Of The Art (SOTA) platform that supports the organization inpatient interactions ingesting hundreds of thousands of phone calls for medical triage and support, and other interactions, surfacing enterprise insights such as anomalies, trending information in drug usage, risk, adverse events, insurance, safety, and general monitoring. The analysis engine, powered by AWS services as a novel cloud deployment for the organization, works to digest patient interactions (>10min audio) and is critical for monitoring chronic diseases, care and progression, and patient events across dozens of markets.
The primary objective from prototype to production was to enable better patient management and analysis, generating insights from raw phone call audio and matching respective unstructured call manager (nurse, assistant, etc.) notes into key text, structured topics, and trends, displayed in a searchable ‘command center’ view. Our success was in developing a custom and innovative system beyond the standalone capabilities of the Nuance Transcription Engine, AWS Transcribe, and commercial NLP services such as AWS Comprehend, by unifying on-premise clinical management software with APIs connecting to the transcription, NLP, and backend while algorithmically merging caller and medical notes metadata.
We developed an AI-based speech-to-text validation of the phone call transcripts, a keyword and entity extraction pipeline to show word and phrase trends over time, a topic model pipeline to show topic frequency trends in patient groups, packaged in a deployable application and UI that visualizes the resulting time-series trends across user-specific and operational data cuts. The solution included developing the relevant database, ETL pipelines, including the data governance procedures, database infrastructure, and schemas for proper storage, validation, archiving, and analysis functions linking the transformed datasets. Designed for hybrid environments, it enables the architecture to function as a high-performance and reliable serverless database and avoids traditional overhead per functional specifications. This allowed for on-demand scaling of our NLP services, seamless export to BI tools capabilities, and capture and archiving of metadata such as raw audio files, pre/post-processed audio records, unstructured notes and actions, derived tagged text data, and associated information including patient demographics, case manager IDs, length of calls, voicemail messages, and other PHI and proprietary metadata and medical information.
Data privacy is maintained via a multi-pronged approach at various points in the pipeline: During data ingestion and metadata merging, all data is cleaned for any Personally Identifiable Information (PII) and only an anonymized ID is kept. Further, to anonymize transcribed audio data, every transcript is scanned and a Named-entity recognition (NER) system is used in conjunction with existing first and last name databases for existing patient names, in addition to confidential operational programs, to mask transcripts before information extraction processing.