Enabling value-based healthcare through automating risk assessment for episode-based care

Period of Performance: 09/01/2017 - 02/28/2018


Phase 1 SBIR

Recipient Firm

Capsicohealth, Inc.
Principal Investigator
Principal Investigator


Project Summary Value-based healthcare implementation relies on understanding risk. 1 Early models, such as Medicare Advantage, use annual measures of risk under a risk adjustment factor (RAF) to offer financial incentive to payers and hospitals to work together. 2 More advanced models, such as bundled payments, target the periods of greatest quality variability, specifically episodes of care such as joint replacement, oncology diagnosis, and cardiac procedures. In these episodes, many types of providers, from hospitals to outpatient physical therapists, need to work together to reduce rates of complication and readmission. Risk levels are used to adjust payment for payer and providers and to determine which patients require additional resources in the hospital, clinic, or home. Unfortunately, existing risk models lack key features needed for episode-based care, which requires both financial alignment and accurate and immediate information to adjust clinical resources for a given case. 3 4 A better model would include all conditions relevant to an episode rather than just chronic conditions, addition of social determinants, and an automated approach to retrieve the information in hours rather than months. Thus, this Small Business Innovation Research (SBIR) Phase I program includes the following Specific Aims: 1. Create the phenotyping components required to define an accurate and comprehensive model of episode-based risk, including: (i) extract clinical and social features from clinical data using natural language processing (NLP), (ii) map concepts including social features to an ontology that will support normalized data use, (iii) build a feature vector for each record that can be used to feed a risk model that accounts for relevant clinical and social risk 2. Validate the phenotyping components using de-identified longitudinal clinical data for 10,000 patients In this research program, Phase I will tackle the most difficult challenges, including leveraging narrative text to recognize time-labeled social and clinical features influencing an episode of care. Success criteria will be accurate recognition of key underlying features that have not been available in risk models to date. Phase II will build upon the validated technology to create an episode-based risk model run on narrative and discrete clinical data and tested against actual patient outcomes. Success criteria will be a validated episode-based risk model to support value-based contracting and value-based clinical care.