OR WAIT null SECS
Currently, diagnosing a patient with systemic lupus erythematosus (SLE) is a complex process that compares potential lupus with other conditions. It can be challenging and delayed by a period of time, which increases patient uncertainty, referrals, healthcare utilization, increased flares, and organ dysfunction. In this study, machine learning (ML) via artificial intelligence tools based on patient data was used to develop an algorithm to help with SLE diagnosis.
Investigators developed an accurate, clinician-friendly algorithm designed to diagnose and treat early systemic lupus erythematosus (SLE) and improve patient outcomes, according to a study in BMJ Journals.1 Currently, diagnosing a patient with SLE is a complex process that compares potential lupus with other conditions. It can be challenging and delayed by a period of time, which increases patient uncertainty, referrals, healthcare utilization, increased flares, and organ dysfunction. In this study, machine learning (ML) via artificial intelligence tools based on patient data was used to develop an algorithm to help with SLE diagnosis.
A study of 802 randomly selected adults with SLE or control rheumatologic diseases that had clinically selected panels of classification criteria: the Systemic Lupus International Collaborating Clinics (SLICC), European League Against Rheumatism (EULAR), and the American College of Rheumatology (ACR) and non-criteria features were analyzed. The model created produced SLE risk probabilities correlating with disease severity and organ damage, which allowed the classification of a validation cohort into diagnostic certainty levels (unlikely, possible, likely, definitive). The overall accuracy was 94.8% for diagnosing SLE and the results for early disease diagnosis were 93.8%. Nephritis had 97.9% diagnosis accuracy, neuropsychiatric was 91.8%, and severe lupus requiring immunosuppressives/biologics was 96.4%. The scoring system used had a 94.2% accuracy.
Data from the Rheumatology Clinics at the University Hospital of Heraklion and the “Attikon” University Hospital, Athens, were used as both centers had established SLE registries and use homogenized, structured forms for collecting clinical characteristics use of treatments and disease outcomes. Patients diagnosed between January 2005 and June 2019 with SLE or rheumatological diseases that were relevant to lupus were eligible for the study. In total, data from 401 patient with SLE and 401 control patients were used to compare the ML models. Diagnostic accuracy was ensured by an external validation cohort of 512 patients with SLE and 143 controls. Patients with early SLE were defined as less than 24 months since diagnosis.
“In clinical practice, physicians can elicit the diagnosis of SLE even in the presence of a few high-yield manifestations such as typical malar rash in an individual with anti-DNA autoantibodies,” investigators stated. “Such decisions reflect a form of human intelligence that develops through clinical experience even with a limited number of patients. Conversely, computational intelligence tools require training on large comprehensive data sets to produce valid results. We used a discovery sample of well-characterized SLE and control patients for unbiased selection of features that contribute most to clinical SLE diagnosis. Patients with SLE with relatively early disease (median duration 4.2 years) and irrespective of the severity of manifestations were included, as compared with developing classification criteria, which typically rely on cases with long-standing disease.”
In the study, 2 approaches were used to create a predictive model for SLE. In the first approach, investigators combined SLE classification criteria with non-criteria features. In the second, a de novo model based on clinical variables of classification criteria and non-criteria features. A 10-fold stratified CV process was conducted, and each fold was used to determine the model performance, while the other folds were used as training data. Sensitivity, specificity, and accuracy were investigated.
The study was limited by its retrospective design and therefore some data may have been missed or underestimated and an early diagnosis would ideally be based on patients with early disease prior to adverse outcomes. Additionally, studies should be used to evaluate and validate these results. “Notwithstanding, our analysis might provide useful insights towards the possible future development of formal SLE diagnostic criteria, a currently unmet need,” investigators concluded. “To this end, establishing a firm diagnosis and treatment plan still remains at the judgement of experienced physicians.”
Adamichou C, Genitsaridi I, Nikolopoulos D, et al. Lupus or not? SLE Risk Probability Index (SLERPI): a simple, clinician-friendly machine learning-based model to assist the diagnosis of systemic lupus erythematosus [published online ahead of print, 2021 Feb 10]. Ann Rheum Dis. 2021;annrheumdis-2020-219069. doi:10.1136/annrheumdis-2020-219069