Effectiveness of artificial intelligence in classification of connective tissue diseases in patients with anti-nuclear antibody (ANA) positivity

Objectives

The study aimed to investigate the classification performance of artificial intelligence (AI) in diagnosing connective tissue diseases(CTD). This was done by analyzing laboratory data, including additional markers, in patients who tested positive for antinuclear antibody(ANA).

Material/Methods

The research included 663 ANA-positive patients. An automated machine learning approach, specifically Auto-Weka, was used to classify these patients based on 75 features, including age, sex, and various laboratory tests.

Results

The Bayes Network achieved the highest overall performance with 93.1 % accuracy, 77.7 % sensitivity, and 96.0 % specificity in the classification of all patients. The most successful models were Locally Weighted Learning for systemic lupus erythematosus(SLE), with an accuracy of 93.4 %; Logistic Model Trees for primary Sjogren's syndrome(pSS), with an accuracy of 91.4 %; AdaBoostM for rheumatoid arthritis(RA), with an accuracy of 95.2 %; and Sequential Minimal Optimization for systemic sclerosis(SSc), with an accuracy of 92.0 %. Sensitivity and specificity rates for SLE, pSS, RA and SSc were found to be 69.4 %, 72.0 %, 78.5 %, 75.3 % and 98.7 %, 96.2 %, 98.9 %, 94.9 %, respectively. The area under the ROC curve in the general distribution of the groups was 95.6 %, the highest value in distinguishing was 99.1 % for RA and the lowest was 85.1 % for SSc. The most predictive markers identified were hematocrit for SLE, anti-SSA for pSS, rheumatoid factor for RA, and anti-centromere B positivity for SSc.

Conclusion

AI models are highly successful in classifying ANA-positive patients with great accuracy. AI-based approaches have the potential to assist clinicians in diagnosing autoimmune diseases by providing more accurate and faster results.