JMIR Publications

ABSTRACT

Background:

Frailty is a multidimensional clinical syndrome characterized by diminished physiologic reserve and increased vulnerability to stressors, thus putting older adults at higher risk of adverse outcomes (e.g., falls, mental and physical disability, hospitalization, mortality) in response to even minor stress events. Frailty can be reversed or at least attenuated if detected early, yet early identification remains challenging in primary care due to time- and resource-intensive assessment methods. Artificial intelligence (AI) offers promise in automating frailty identification at the point of care. Natural Language Processing (NLP) is particularly valuable for extracting frailty indicators from rich text data stored in electronic health records, but its limited interpretability has prompted growing interest in augmenting the NLP processes with the use of explainable AI (XAI) techniques. Although NLP and XAI methods have been applied for chronic disease identification, their use for frailty identification has not yet been systematically examined.

Objective:

This scoping review aimed to synthesize current evidence on the use of NLP and XAI methods for automating frailty identification in older adults.

Methods:

Peer-reviewed studies published in English between January 2015 and November 2025 were eligible if they applied AI, NLP, or XAI methods to identify frailty in adults aged ≥50 years using real-world health data from OECD or OECD-partner countries. Searches were performed in PubMed and Google Scholar and supplemented by screening bibliographies of identified studies. Data were extracted using a standardized form that captured study characteristics, sample size, data sources, and specific aspects of the AI models, and NLP and XAI methods used.

Results:

We identified 24 studies that satisfied the eligibility criteria. While all studies used AI approaches to identify frailty, only six used neural network-based models. Logistic regression was the most frequently used AI method (n=14), and only one study employed Bidirectional Encoder Representations from Transformers (BERT). Seven studies relied on both structured and unstructured data, two relied exclusively on structured data only, and the rest relied exclusively on unstructured data. Seven studies used NLP methods, seven used XAI methods, and only one integrated both. Only two studies reported deploying their models in real clinical settings.

Conclusions:

AI-based approaches show promise for automating frailty identification, yet current applications remain limited by reliance on traditional machine learning models, underuse of NLP and XAI methods, and very little real-world deployment. Future work should focus on developing explainable NLP models, facilitating access to large volumes of unstructured data, and developing standardized frameworks for the systematic evaluation of NLP and XAI methods. Coordinated efforts across clinical, technical, and regulatory domains are essential to develop scalable, transparent, and clinically meaningful AI systems for frailty identification.