Harnessing Natural Language Processing to Measure Person-Centered Care in Behavioral Health Settings

Background and Purpose: While receiving widespread endorsement both in medical and behavioral health care settings, operationalizing and documenting person-centered care has been a challenge. Most Person-Centered Care (PCC) measures rely on self-report which can be unreliable, prompting the need for objective and feasible ways to document PCC. Clinical narrative notes provide a rich, but relatively untapped, source of unstructured data that can be an important indicator of PCC. Natural language processing (NLP) is an artificial intelligence method that has the capability to convert clinical narratives embedded within the EHR into structured data. The purpose of this study was to utilize NLP to develop a measure of PCC for behavioral health care settings.

Methods: We utilized a supervised ontology-based information extraction approach, an NLP method which has been widely used in medical settings. The dataset (or corpus) is clinic visit notes completed by providers in a community mental health center. These notes are narrative documents completed when the provider has face-to-face in person contact with a service user to develop, revise or document progress on the service plan. Following the Noy & McGuiness ontology development 101 procedure, we developed an ontology to capture the domains of person-centered care. Ontology structures were developed using Protégé software and frequency analysis and validation were performed using Python 3.0.

Results: The ontology development process involved the following steps 1) we grounded the domains of PCC in the items of the Person-Centered Care Planning Assessment Measure and the Person-Centered Care Planning intervention, and the broader person-centered care literature, 2) we enumerated the terms under the high-level concepts according to the sources, 3) we determined hierarchies and relationships between PCC concepts and language, 4) we supplemented the list of terms that could belong to each concept through term frequency-inverse document frequency analysis and bag of bigrams model of the corpus of clinical visit notes, 5) we mapped identified terms to their concepts using the ontology, and 6) we validated the ontology by examining the degree of representability on our collected clinical visit notes.

Conclusions and Implications: This study demonstrated that complex and abstract practices such as PCC can be amenable to NLP methods to generate objective measures. This deductive ontological approach has the potential to provide a scalable method that mines clinically rich unstructured data documenting service processes to measure aspects of service quality in behavioral health settings, and can be utilized both in research and quality improvement efforts.