Clinical documentation · Production

Clario automates clinical research COA interview analysis using generative AI on AWS

The problem

Traditional COA interview quality evaluation required time-consuming, logistically challenging reviews of audio-video recordings in near real time, with variability between expert reviewers, poor assessment technique, and other noise factors that could lead to unreliable results and study failure.

Workflow diagram · grounded in source
1
Interview recordings uploaded to S3
trigger
“The COA interview recordings (audio and video files) from the interviews are collected on premises (1) using a recording application. The files are uploaded using AWS Direct Connect with encryption in transit to Amazon Simple Storage Ser…”
2
Speaker diarization on SageMaker
ai_action
“extracts the audio and identifies speech segments of unique speakers using a custom speaker diarization model on Amazon SageMaker”
3
Multi-lingual transcription via Whisper
ai_action
“Clario uses the Whisper model from the Amazon Bedrock Marketplace (5) to generate near real-time transcriptions of the COA interview recordings. The transcriptions are then annotated with speaker information and timecodes”
4
Vectorization and storage in OpenSearch
integration
“vectorized using an embedding model (Amazon Titan Text Embeddings v2 model) and stored into Amazon OpenSearch (7) for semantic retrieval”
5
Graph-based agent COA review
ai_action
“Clario's AI Orchestration Engine executes a graph-based agent system running on Amazon Elastic Kubernetes Service (Amazon EKS) (3) for automated COA interview review. The agent implements a multi-step workflow that: (1) retrieves the ass…”
6
LLM classification and quality evaluation
ai_action
“The agent uses advanced large language models (LLMs), such as Anthropic Claude 3.7 Sonnet from Amazon Bedrock (6), to classify the speech segment as interviewer or participant, and to determine if each interview turn meets the interview …”
7
AI output validation against source
validation
“Clario to implement a validation system where AI outputs are cross-checked against source documents for factual accuracy before human review”
8
Review compiled and persisted to RDS
output
“Clario's AI Orchestration Engine then compiles the overall review of the interview and persists the information in Amazon Relational Database Service (Amazon RDS)”
Reported outcome

Clario's AI-powered solution shows potential to decrease manual review effort by over 90%, achieve up to 100% data coverage through automated review, and shorten central review turnaround time from weeks to hours.

Reported metrics
Manual review effort reductionover 90%
Data coverage via automated reviewup to 100%
Central review turnaround timefrom weeks to hours
Reported stack
Amazon BedrockAmazon SageMakerAmazon Titan Text Embeddings v2Amazon OpenSearchAmazon EKSAnthropic Claude 3.7 SonnetAmazon RDSAmazon API GatewayAWS Lambda
Source
https://aws.amazon.com/blogs/machine-learning/how-clario-automates-clinical-research-analysis-using-generative-ai-on-aws?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Clario's AI-powered solution shows potential to decrease manual review effort by over 90%, achieve up to 100% data coverage through automated review, and shorten central review turnaround time from weeks to hours.

What tools did this team use?

Amazon Bedrock, Amazon SageMaker, Amazon Titan Text Embeddings v2, Amazon OpenSearch, Amazon EKS, Anthropic Claude 3.7 Sonnet, Amazon RDS, Amazon API Gateway, AWS Lambda.

What results were reported?

Manual review effort reduction: over 90%; Data coverage via automated review: up to 100%; Central review turnaround time: from weeks to hours (source-reported, not independently verified).

How is this clinical documentation AI workflow structured?

Interview recordings uploaded to S3 → Speaker diarization on SageMaker → Multi-lingual transcription via Whisper → Vectorization and storage in OpenSearch → Graph-based agent COA review → LLM classification and quality evaluation → AI output validation against source → Review compiled and persisted to RDS.

Related clinical documentation cases