Options
June 10, 2024
Journal Article
Title
POS0273 A machine learning algorithm based on automatically extracted image features and clinical data on morning stiffness can distinguish psoriatic arthritis patients from controls in clinical routine care
Abstract
Background: Diagnosis and treatment of psoriatic arthritis (PsA) is often delayed due to a lack of clear diagnostic criteria and limited resources for referral to a rheumatologist, including a high number of inappropriate referrals. Artificial intelligence (AI) and natural language processing (NLP) methods provide algorithms for learning systems to recognize disease-related terms and classify clinical phenotypes using large datasets, which may support early identification of patients with a suspected diagnosis of PsA. In previous work [1], we presented an approach using NLP to identify morning stiffness (MST) as a predictor to discriminate patients with diagnosed PsA and axSpA from controls suspected of having PsA or axSpA. With the same aim, we extended the panel of predictors by combining image features extracted from hand and foot radiographs for the PsA cohort.
Objectives: AI and NLP methods are used to identify patients with typical radiographic features (RGF) of inflammation as potential discriminators that can be automatically recognized to support referral.
Methods: A multicentre observational pilot study recruited patients seen by a rheumatologist with a suspected diagnosis of PsA (n=36; classified PsA n=28 according to CASPAR, No PsA=8) from the referring primary care provider. All clinical examination data and findings were collected and adjudicated by rheumatologists, focusing on criteria for the diagnosis of PsA. Unstructured text from the medical history was used to extract diagnostic features including MST. The information extraction algorithms used NLP models to detect expert-curated MST keywords (inclusion of the textual context) and use the results as structured data. In addition, X-ray data was automatically extracted, and bone segmentation was performed, followed by the calculation of “diagnostic parameters”. For bone segmentation, features such as joint/bone distance etc. (Figure 1) were selected as diagnostic parameters. In total, 504 features were identified for each hand and reduced to 159 features without critical loss of quality. To finally merge the tabular clinical data set with the image features, different ML model architectures (e. g. logistic regression) were tested and compared via 5-fold cross-validation.
Results: The combination of RGF and structured data is summarized in Table 1. With the presented approach, we achieved the best results with an F1 score of 0.89 (±0.07) for the automated detection of MST when using image features extracted from non-reduced images in combination with tabular data and logistic regression classifier. A reduced number of features increases the explanatory power.
Conclusion: The presented combination of tabular clinical data on MST and extracted image features shows promising results for the detection of classified PsA and may in the future assist physicians. A study is planned not only to confirm the results but also to refine the ML models developed and specify disease attributes.
Objectives: AI and NLP methods are used to identify patients with typical radiographic features (RGF) of inflammation as potential discriminators that can be automatically recognized to support referral.
Methods: A multicentre observational pilot study recruited patients seen by a rheumatologist with a suspected diagnosis of PsA (n=36; classified PsA n=28 according to CASPAR, No PsA=8) from the referring primary care provider. All clinical examination data and findings were collected and adjudicated by rheumatologists, focusing on criteria for the diagnosis of PsA. Unstructured text from the medical history was used to extract diagnostic features including MST. The information extraction algorithms used NLP models to detect expert-curated MST keywords (inclusion of the textual context) and use the results as structured data. In addition, X-ray data was automatically extracted, and bone segmentation was performed, followed by the calculation of “diagnostic parameters”. For bone segmentation, features such as joint/bone distance etc. (Figure 1) were selected as diagnostic parameters. In total, 504 features were identified for each hand and reduced to 159 features without critical loss of quality. To finally merge the tabular clinical data set with the image features, different ML model architectures (e. g. logistic regression) were tested and compared via 5-fold cross-validation.
Results: The combination of RGF and structured data is summarized in Table 1. With the presented approach, we achieved the best results with an F1 score of 0.89 (±0.07) for the automated detection of MST when using image features extracted from non-reduced images in combination with tabular data and logistic regression classifier. A reduced number of features increases the explanatory power.
Conclusion: The presented combination of tabular clinical data on MST and extracted image features shows promising results for the detection of classified PsA and may in the future assist physicians. A study is planned not only to confirm the results but also to refine the ML models developed and specify disease attributes.
Author(s)
Barton, Anne
Division of Musculoskeletal & Dermatological Sciences, University of Manchester, Manchester, United Kingdom
Coates, Laura
University of Oxford, Nuffield Department of Orthopedics, Rheumatology and Musculoskeletal Disorders
Pennington, Stephen
University College Dublin, School of Medicine, St Vincent’s Hospital Elm Park Dublin
Fitzgerald, Oliver
University College Dublin, School of Medicine, St Vincent’s Hospital Elm Park Dublin
Conference