85. Development Of An AI-Based Predictive Model For Septic Wrist And A Risk Assessment Tool

Abstract

Purpose: The existing diagnostic criteria for septic wrist are non-specific, subjecting patients with non-infectious etiologies to surgical morbidity. This study aimed to create a septic wrist AI prediction model and develop a score-based risk assessment tool. Methods: An IRB-approved retrospective review was conducted on patients with a presumed septic wrist diagnosis (2003 - 2022). Patients were excluded if neither arthrocentesis nor open drainage was performed. Kruskal Wallis algorithm was employed to identify potential predictors of septic wrist based on factors including comorbidities (autoimmune diseases, immunosuppression, prior crystalline arthropathy, prior septic arthritis, IVDU, smoking), penetrating trauma, fever, multi-joint involvement, inflammatory markers (ESR/CRP/WBC), serum uric acid, blood cultures, imaging, and synovial fluid analysis. Subsequently, Naïve Bayes classifier was utilized to populate a prediction model based on the identified predictors. Concurrently, an independent score-based risk assessment tool was developed using traditional multivariate analyses. In this tool, each predictor received a standardized risk score of 1, and the optimal cutoff for classifying patients with septic wrist was established. Performance of both methods were evaluated with ROC curve. Results: 205 (70 females and 135 males) patients were included with an average age of 60 ± 16 years. The median length of hospitalization was 6 [8] days and follow-up duration 1 [3] months. 95 (46.3%) patients had septic wrist confirmed with Gram stain/culture, 79 (38.5%) patients received alternative diagnoses (crystalline arthropathy, tenosynovitis, cellulitis, abscesses, ligament injury, and arthritic flare), and 31 (15.1%) patients had undetermined diagnoses. Concomitant septic wrist and crystalline arthropathy was identified in 11 patients (5%). The optimized Naïve Bayes prediction model included 7 predictors (no synovial crystals, positive blood culture, multi-joint involvement, age, history of septic arthritis, IVDU, and penetrating trauma) with a sensitivity of 89.5%, specificity of 80%, and an AUC of 0.89. The score-based risk tool comprised 4 predictors (negative synovial crystals, positive blood culture, negative history of crystalline arthropathy, and multi-joint involvement), resulting in a risk score ranging from 0 to 4. Classifying patients with septic wrist at a score of ≥2 yielded an optimized sensitivity of 64%, specificity of 84%, and AUC of 0.78. Conclusion: Both the AI prediction model and the traditional statistical score-based risk assessment tool can offer valuable support for clinical decision-making in cases of suspected septic wrist. The most significant predictors for septic wrist include multi-joint involvement, absence of synovial crystals, and positive blood culture.

Ähnliche Arbeiten