Intelligent HCC risk scoring with NLP to identify high-risk patients

Introduction

A global healthcare software and services provider sought to optimize their HCC (Hierarchical Condition Category) scoring process. Their objective was to automate the extraction of critical clinical insights from unstructured patient records, including scanned and handwritten documents, to identify high-risk patients and support more accurate claims reimbursement. They worked with CitiusTech to develop the solution.

The Challenges
Unlocking critical risk data from complex, unstructured sources

The client faced several hurdles in their HCC scoring workflow:

Unstructured inputs: Clinical diagnoses buried in mixed-format records, including scanned PDFs and handwritten notes.
Manual overhead: Labor-intensive processes slowed down risk scoring and claim reviews.

They required a scalable, intelligent solution to extract, interpret, and structure this data for accurate HCC scoring.

The Solution
Intelligent automation pipeline combining NLP, OCR, and deep learning

Built a modular pipeline leveraging advanced NLP and Optical Character Recognition (OCR) technologies to automate clinical data abstraction and HCC score computation. Key components included:

Input handling: Consolidated PDFs of health records (typed and handwritten).
YOLOv3 with transfer learning: Detected and interpreted handwritten objects, including signatures.
Section classification models: Classified each section header in a clinical note.
Document classification engine: Categorized each PDF page into non-clinical, encounter start, intermediate, or end.
Entity extraction: Captured provider name, credentials, encounter details, electronic/handwritten signatures, and calculated confidence score.
CMS risk API integration: Populated HCC scores using standardized CMS data.
Disease interaction engine: Mapped co-morbidities and associations to flag patients with high-severity conditions.

Fig 1: Solution Highlight

Value Delivered
Measurable improvements in risk identification and operational efficiency

60% Reduction in claim review time	Automated pipelines accelerated risk analysis.
Improved claim reimbursements	Accurate HCC scoring led to better coding and documentation.
High model performance
1. Handwritten signature extraction	Precision: 0.71 Accuracy: 0.86
2. Section classification	Accuracy: 0.93–0.99 F1 Score: 0.42–0.93 AUC: 0.91–0.99
3. Entity extraction	Provider name accuracy: 0.95 E-signature accuracy: 0.89

Enhancing chronic care outcomes through intelligent data abstraction

This solution offers a replicable model for healthcare and life sciences organizations, aiming to:

Automate and standardize clinical abstraction
Improve documentation accuracy for better reimbursement
Quickly flag high-risk patients to support care coordination

Reimagine Discharge Management with Gen AI

Explore how leading hospitals are saving hours per patient using Gen AI-enabled task evaluation.

Power intelligent care journeys