Risk index provides 'clinical aid' to differentiate lupus, other rheumatic diseases
Click Here to Manage Email Alerts
Researchers have used machine learning to develop and validate a new model that can produce individualized risk probabilities for systemic lupus erythematosus, according to findings published in the Annals of the Rheumatic Diseases.
The model, called SLE Risk Probability Index (SLERPI), demonstrates an excellent combination of sensitivity and specificity for SLE against competing rheumatic diseases when treated as a binary — SLE or not SLE — the researchers wrote. In addition, the logistic regression model can be converted into a simple scoring system for clinical and serological features.
“SLE can be challenging to diagnose especially at early stages,” George K. Bertsias, MD, PhD, of the University of Crete School of Medicine, in Heraklion, Greece, told Healio Rheumatology. “This study used sophisticated machine learning algorithms in large, well-characterized patient cohorts in order to simulate clinical diagnostic thinking. Through this approach, we developed and validated a simple index — SLERPI — which provides individualized risk probabilities for the presence of SLE against common mimicking rheumatological diseases.”
To develop an algorithm that can help providers diagnosis SLE, Bertsias and colleagues applied machine learning to well-characterized patient data sets. Working with rheumatology clinic data from the University Hospital of Heraklion and the Attikon University Hospital, in Athens, the researchers used a randomly selected discovery cohort of 401 patients with SLE and 401 controls with miscellaneous rheumatic diseases to construct, train and compare the machine learning models.
In addition, they used an external validation cohort of 512 consecutively registered patients with SLE and 143 controls to give an unbiased estimate of the diagnostic accuracy of the best model. Feature selection and model construction were completed with Random Forests and Least Absolute Shrinkage and Selection Operator-logistic regression (LASSO-LR). The researchers then used the validation cohort to test the best model in 10-fold cross-validation.
According to the researchers, the best performing model was the novel LASSO-LR, which included 14 variably weighed features with thrombocytopenia/hemolytic anemia, malar/maculopapular rash, proteinuria, low C3 and C4, antinuclear antibodies and immunologic disorder being the strongest predictors of SLE. Depending on this combination of features, the model produced SLE risk probabilities that positively correlated with disease severity and organ damage. This allowed for the unbiased classification of a validation cohort into diagnostic certainty levels — unlikely, possible, likely or definitive SLE — based on the likelihood of this disease against other diagnoses.
In addition, operating the model as a binary — ie, lupus or not-lupus — the researchers reported 94.8% accuracy for identifying SLE, and a high sensitivity for early disease, at 93.8%; nephritis, at 97.9%; neuropsychiatric disease, at 91.8%; and severe lupus requiring immunosuppressives/biologics, at 96.4%. Bertsias and colleagues then converted this into a scoring system, in which a score of more than 7 corresponded to 94.2% accuracy.
“In routine practice, the index can be used to classify individuals according to varying diagnostic certainty levels — ie, ranging from ‘unlikely’ to ‘definitive’ SLE — therefore serving as a clinical aid to physicians,” Bertsias said. “As lupus tends to develop in stepwise fashion, the SLERPI can also be used for the prospective monitoring of individuals with feature suggestive of a connective tissue disease. Finally, this work represents a step forward to establishing SLE diagnostic criteria, which are currently missing.”
Collapse
Read more about
Click Here to Manage Email Alerts