Advancing Patient Experience and Business Operations Analytics with Generative AI in Healthcare

huron · Healthcare · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Healthcare analytics

Huron needed to analyze unstructured patient-experience and revenue-cycle notes quickly enough to support operational interventions.

Constraints:

Healthcare workflows
Unstructured notes
Traditional survey lag
Accuracy requirements

Architecture

Huron's healthcare analytics system turns unstructured patient-rounding and business-operations notes into sentiment and opportunity signals inside an AWS data stack. EventBridge schedules a Step Function, the workflow invokes Redshift stored procedures, Nova Micro is registered as a Redshift-callable model through Bedrock, prompts classify each question/answer into one of four sentiments, and client business-operation data flows through S3, Glue, dbt, and Redshift before BI aggregation. The same pattern supports smart rounds, HCAHPS intervention, net-revenue opportunity detection, denial-trend analysis, and staff coaching.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  rounding-platform[Rounding platform] --> sentiment-workflow[Sentiment workflow]
  client-raw-data[Client raw data] --> s3-storage[S3 storage]
  s3-storage --> glue-catalog[Glue catalog]
  glue-catalog --> redshift-and-dbt[Redshift and dbt]
  eventbridge-schedule[EventBridge schedule] --> sentiment-workflow
  sentiment-workflow --> redshift-and-dbt
  sentiment-workflow --> nova-micro-model[Nova Micro model]
  nova-micro-model --> sentiment-outputs[Sentiment outputs]
  sentiment-outputs --> bi-analytics[BI analytics]
  redshift-and-dbt --> bi-analytics
  bi-analytics --> operations-teams[Operations teams]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Rounding platform`	Source of manually entered patient-rounding questions and answers.	`Sentiment workflow`
`Client raw data`	Business-operations notes, flat files, and client data sources.	`S3 storage`
`S3 storage`	Storage landing zone for raw client data.	`Glue catalog`
`Glue catalog`	Catalog layer for raw and transformed business-operations data.	`Redshift and dbt`
`EventBridge schedule`	Daily trigger for the sentiment workflow.	`Sentiment workflow`
`Sentiment workflow`	Step Function and stored-procedure workflow that invokes model scoring.	`EventBridge schedule`, `Redshift and dbt`, `Nova Micro model`
`Redshift and dbt`	Warehouse and transformation layer for curated analytics tables and model invocation.	`Glue catalog`, `Sentiment workflow`, `BI analytics`
`Nova Micro model`	Bedrock-hosted model used to classify question/answer sentiment.	`Sentiment workflow`, `Sentiment outputs`
`Sentiment outputs`	Four-category sentiment labels and themes for notes and operations data.	`Nova Micro model`, `BI analytics`
`BI analytics`	QuickSight and analytics layer for themes, opportunities, and operational interventions.	`Redshift and dbt`, `Sentiment outputs`
`Operations teams`	Healthcare users who act on service recovery, revenue, denial, and coaching insights.	`BI analytics`

Data flow

Manual question-and-answer notes from rounding and raw business-operations data are stored and curated in AWS. EventBridge triggers a Step Function, Redshift invokes stored procedures that call Nova Micro through Bedrock, the prompt returns one of four sentiments per question/answer, outputs are stored and aggregated with BI tools, and operations teams use them for service recovery, revenue-cycle, and coaching actions.

3 supporting quotes

“The EventBridge is just a glorified schedule that triggers the step function Huro-Sentiment-Daily, which invokes Redshift's storage proc.”transcript, stated

“We register the Nova Micro model which shows up in Redshift as a function that can be called. If you are familiar with the MMM model in general, the concept of the proc, the prompt within the model passes the question and answer text, which instructs the model”transcript, stated

“which instructs the model to return one of four sentiments based on each answer and question. - By way of comparison, I think the business operations architecture functions in much the same flow. We are receiving raw data from our clients sometimes in the form”transcript, stated

Why this architecture

Huron used an AWS warehouse-native architecture because its inputs already lived as operational data and free-text notes, while the business needed faster intervention than delayed HCAHPS surveys. Registering the model as a Redshift function let sentiment scoring run close to curated analytics data and feed existing healthcare BI workflows.

3 supporting quotes

“As stated, HCAHPS funding takes about three to four months to get back. So, with us being able to detect poor patient experience earlier on, we'll hopefully increase those HCAHPS scores that the hospitals are seeing, which will in return impact their funding.”transcript, stated

“To capitalize off of this new data that we will be receiving, moving towards creating smart rounds, using real-time sentiment analysis to identify patients at risk of bad experiences to provide target service recovery interventions is going to be key.”transcript, stated

“We bring that and store that in S3, move it to tables in Glue catalog, and then we begin to really curate the data through transformations in Redshift powered by dbt.”transcript, stated

Models used: Amazon Bedrock Nova LLM · Nova Micro

Scale

Sentiment accuracy improved from about 75% to 90%.
“We started at about 75% and have worked up to 90%. We understand, especially within healthcare, that we need to get to as close to 100% as possible.”transcript, stated
The business-operations workflow reviews 10,000 notes per week on an automated cadence.
“So, we're at 10,000 notes that are reviewed on an automated cadence per week now while we maintain around a 90% accuracy.”transcript, stated

Alternatives considered (roads not taken)

Wait for HCAHPS survey feedback

Rejected because: Three-to-four-month lag loses the window for service recovery.

“The downfall to this is that it does take three to four months before the data is received back and by then, missed opportunities, early intervention to issues has been lost. - And I think we can move to our architecture slides now.”transcript, stated

Manual review of free-text notes

Rejected because: Automated sentiment scoring scales to weekly note volumes while preserving around 90% accuracy.

“So, we're at 10,000 notes that are reviewed on an automated cadence per week now while we maintain around a 90% accuracy. We'll be able to expand that greatly as we move to broader sections of our client base.”transcript, stated

Patterns

`batch-offline-processing` (analytics)

The system processes large volumes of notes for healthcare operations analytics.

“tes. The solution achieved 90% accuracy in sentiment classification (up from 75% initially) and now processes over 10,000 notes per week automatically, enabling real-time identification of patient dissatisfaction, revenue opportunities, and staff coaching needs that directly impact hospital funding”summary, stated

`structured-output-contract` (analytics)

Sentiment classification converts unstructured interactions into operational categories.

“responses, staff interactions, and financial operation notes. The solution achieved 90% accuracy in sentiment classification (up from 75% initially) and now processes over 10,000 notes per week automatically, enabling real-time identification of patient dissatisfaction, revenue opportunities, and st”summary, stated

Tradeoffs

Prioritize real-time service recovery signals over waiting for formal survey cycles.

Chose Smart rounds with real-time sentiment analysis over Delayed HCAHPS-only analysis after discharge.

Rationale: Earlier detection allowed staff to prioritize patients at risk of bad experiences before the feedback window closed.

“moving towards creating smart rounds, using real-time sentiment analysis to identify patients at risk of bad experiences to provide target service recovery interventions is going to be key.”transcript, stated

How quickly to surface patient-experience risk.

Chose real-time smart-round sentiment analysis over HCAHPS-only delayed feedback.

Rationale: Earlier signal lets teams intervene while the patient experience can still be recovered.

“detect poor patient experience earlier on”transcript, stated

Gotchas (what bit them)

Traditional patient-experience analytics arrived too late for service recovery.

Root cause: HCAHPS survey feedback lagged the patient encounter by three to four months.

Resolution: Huron moved toward real-time sentiment analysis on rounding and staff-interaction notes.

“it does take three to four months before the data is received back and by then, missed opportunities, early intervention to issues has been lost.”transcript, stated

Evaluation

Huron measured sentiment classification accuracy and note-processing throughput.

Metrics: accuracy · notes per week

“automated cadence per week now while we maintain around a 90% accuracy. We'll be able to expand that greatly”transcript, stated

Outcomes

The solution reached roughly 90% sentiment accuracy and automated weekly note processing.

The summary reports real-time identification of dissatisfaction and revenue opportunities.