Researchers use machine learning to create a clinical decision support model that helps healthcare providers predict the diagnosis of hidradenitis suppurativa (HS) and distinguish it from other skin diseases that mimic its symptoms. Developed.1
Machine Learning | Image credit: © WrightStudio – Stock.adobe.com
According to a study published in Frontier of medical technologyThis model could help recognize HS more quickly and accurately, potentially reducing diagnostic delays and costs to the healthcare system.1 To refine and optimize the performance of the model, we recommend further validation through external datasets and clinical practice testing, compared to dermatologist diagnoses.
HS is a chronic inflammatory follicular skin disease characterized by painful intercutaneous lesions that can cause odor, drainage, and disfigurement, leading to a psychosocial burden and worsening quality of life for patients. It leads to Its prevalence varies worldwide, but it is more common among women, smokers, and patients with metabolic syndrome. Furthermore, American studies show that black and biracial patients are two to three times more likely to experience HS than white patients.2
Diagnosis relies on clinical criteria and early detection is important for better management.1 Misdiagnosis or underdiagnosis can lead to prolonged suffering and increased medical costs. Machine learning is increasingly being used to aid in disease recognition, including HS. Applications to electronic medical records and claims databases have also successfully identified a variety of conditions, including depression, ankylosing spondylitis, cardiomyopathy, dementia, and hepatitis C.
The researchers trained and tested their machine learning model using datasets from the IBM MarketScan Research database from 2000 to 2018, and used data from 2018 and 2019 to validate the model. This database contained adjudicated medical and pharmaceutical reimbursement claims for more than 225 million of his patients enrolled in commercial, Medicare, and Medicaid health plans across the United States.
Six single machine learning algorithms and two ensemble methods were considered, and the final model was selected based on performance measurements and consultation with dermatologists. Performance metrics such as area under the curve, sensitivity, precision, and precision were used to evaluate and select the best model, and an accuracy/precision threshold of 0.7 was determined to be satisfactory.
Of the 411,061 HS patients identified from January 2000 to March 2018, 55,989 were included in the study. In addition, 278,483 patients with confirmed abscess and 1,431,524 patients with confirmed cellulite were included as controls.
Key results reveal that high-performance machine learning models for predicting HS diagnoses can be built using claims data, with top models achieving up to 65%-73% diagnostic accuracy and 81%- We achieved an area under the curve of 82%. The model trained to distinguish between HS and cellulitis performed better than the model trained on abscesses, likely due to the similarity between abscesses and HS lesions. The top three models identified were AdaBoost, LightGBM, and MaxVoting, with age, gender, and certain risk factors being strong predictors. Additionally, diagnostic characteristics and specific comorbidity diagnoses were significant predictors across different algorithms and cohorts.
Sensitivity analysis and validation results show that shorter time frames around the index date yield comparable performance metrics for predicting HS, demonstrating that shorter data periods are reliable for model development in claims analysis. Suggests. Validation results show consistent performance among the top three models, predicting 64%–69% of true she-HS patients, with models 1 and 2 showing better performance.
Exploratory applications revealed significant underdiagnosis of HS in patients with abscess or cellulitis, which varied by metropolitan statistical area and model used. This suggests that implementing machine learning models could help health systems identify undiagnosed HS patients for further evaluation and research.
This study addresses the limitations of generalizability, data structure requirements for model application, potential algorithmic variation across populations, and contextual factors such as addressing medical coding errors and temporal relationships between patient claims. Areas for model improvement, such as consideration of
References
1. Kirby J, Kim K, Zivković M, et al. Uncovering the burden of misdiagnosis and underdiagnosis of hidradenitis suppurativa: a machine learning approach. front med technology. Published online March 25, 2024. doi:10.3389/fmedt.2024.1200400
2. Garg A, Kirby JS, Lavian J, Lin G, Strunk A. Sex- and age-adjusted population analysis of prevalence estimates of hidradenitis suppurativa in the United States. JAMA Dermatol. 2017;153(8):760-764. doi:10.1001/jamadermatol.2017.0201
