Automated Coding of Under-Studied Medical Concept Domains: Linking Physical Activity Reports to the International Classification of Functioning, Disability, and Health

Abstract

Linking clinical narratives to standardized vocabularies and coding systems is a key component of unlocking the information in medical text for analysis. However, many domains of medical concepts, such as functional outcomes and social determinants of health, lack well-developed terminologies that can support effective coding of medical text. We present a framework for developing natural language processing (NLP) technologies for automated coding of medical information in under-studied domains, and demonstrate its applicability through a case study on physical mobility function. Mobility function is a component of many health measures, from post-acute care and surgical outcomes to chronic frailty and disability, and is represented as one domain of human activity in the International Classification of Functioning, Disability, and Health (ICF). However, mobility and other types of functional activity remain under-studied in the medical informatics literature, and neither the ICF nor commonly-used medical terminologies capture functional status terminology in practice. We investigated two data-driven paradigms, classification and candidate selection, to link narrative observations of mobility status to standardized ICF codes, using a dataset of clinical narratives from physical therapy encounters. Recent advances in language modeling and word embedding were used as features for established machine learning models and a novel deep learning approach, achieving a macro-averaged F-1 score of 84% on linking mobility activity reports to ICF codes. Both classification and candidate selection approaches present distinct strengths for automated coding in under-studied domains, and we highlight that the combination of (i) a small annotated data set; (ii) expert definitions of codes of interest; and (iii) a representative text corpus is sufficient to produce high-performing automated coding systems. This research has implications for continued development of language technologies to analyze functional status information, and the ongoing growth of NLP tools for a variety of specialized applications in clinical care and research.

Linking Free Text Documentation of Functioning and Disability to the ICF with Natural Language Processing

“…Classification models consistently outperformed candidate selection (p=0.041 for Mobility without Action oracle; p≪0.001 for Mobility with Action oracle and both settings of Self-Care/Domestic Life). This is consistent with our prior findings of comparable or slightly lower performance for our candidate selection model on Mobility data from physical therapy encounters (16). The Action oracle significantly (p≪0.001) improved performance in all cases, clearly demonstrating the value of building NLP systems to extract the Action components of activity mentions.…”

“…Notes are associated with admissions to ICU units of three academic hospitals in Boston between 2001-2012, and are commonly used for language modeling in clinical NLP research. • NIHCC: Over 63,000 free text notes from 10 years of Physical Therapy and Occupational Therapy encounters in the Rehabilitation Medicine Department of the NIH Clinical Center, collected and used for calculating word embedding features in our previous work (16). • SSA: Over 65,000 free text notes associated with disability claims processed by SSA within a five-year period (as described in the "SSA document collection for language modeling" section above).…”

“…In classification approaches, a mathematical representation is calculated for each activity mention, and a model is trained to assign an ICF code to each Action component based on its similarity to previously-observed samples labeled with each ICF code. We adopted the best-performing classification model from our prior work (16), of a Support Vector Machine (28) using word embedding features as input. Given an input activity mention, we calculated its embedding features in one of four ways:…”

“…For the added Therapy and Manage medication labels, we used the names of the corresponding SNOMED CT codes ("Ability to manage medication" and "Compliance behavior to therapeutic regimen", respectively). Further details of the model are presented in (16). Following our prior work, we used a 3-layer DNN with hidden layer size 300 when using static embedding features without the Action oracle, a 3-layer DNN with layer size 600 when using static embeddings with the Action oracle (to match the dimensionality of the concatenated activity mention and Action component embeddings), and a 1-layer DNN with layer size 768 when using BERT embedding features (for which vector dimensionality does not change with the Action oracle).…”

Translational NLP: A New Paradigm and General Principles for Natural Language Processing Research

“…• Newman-Griffis and Fosler-Lussier (2019) developed a flexible method for identifying sparse health information that is syntactically complex (challenging Data Characteristics). • Newman-Griffis and Fosler-Lussier (2021) compared the Task Paradigms of classification and candidate selection paradigms for medical coding in a new domain.…”