Abstract
BACKGROUND: Aspergillus LDBio ICT IgG/IgM lateral flow assay (LFA) is a cheap alternative to Bordier ELISA assay (BEA) for detection of fungal invasion in chronic pulmonary aspergillosis (CPA) patients in resource-limited settings. However, the sensitivity of LFA is inferior to BEA in the diagnosis of CPA and contributed to high misdiagnosis of CPA in low-income countries. [Figure: see text] [Figure: see text] METHODS: We trained causal machine learning (CML) algorithm using symptoms, chest x-ray features, and IgG of LFA for diagnosis of CPA. This multi-center cross-sectional study included 386 pulmonary tuberculosis (PTB) patients, age 16 - 82 years, 214 (55%) female, 205 (53.1%) HIV positive, 138 (35.8%) post-TB, and 248 (64.2%) retreatments, with 97 (25.1%) having CPA. The BEA was positive in 98 (25.4%) cases while LFA was positive in 71 (18.4%) cases. CML is a dual logistic regression models (LR). The first model is a multivariable LR with symptoms and chest x-ray features as predictors. The second model is a simple LR having IgG of LFA as independent predictor. We also trained traditional multivariable LR (ML) with symptoms, chest x-ray features, and IgG of LFA as predictors of CPA. We used cross validation of 10-fold and partitioned data into 80:20 for training and testing samples. CML was compared to baseline model (BM) (physicians manually combining symptoms + chest x-ray findings + LFA) and traditional ML. CML, BM, and ML performance were assessed using the area under the receiver operating characteristic curve (AUC), accuracy, precision, sensitivity, specificity, kappa, and F1-score. [Figure: see text] RESULTS: The sensitivity of CML was 100% and superior to ML (84%) and BM (68%). CML [Accuracy = 95%, AUC = 0.989] was substantially higher than BM [Accuracy = 91%, AUC = 0.833]. The Cohen' s Kappa of CML (0.91) surpassed BM (0.79). The specificities > 90% for BM, ML, and CML. CONCLUSION: CML algorithm outperformed BM and reduced misdiagnosis of CPA among PTB patients. CML algorithms will aid physicians in optimal diagnosis of CPA in resource-challenge environments. DISCLOSURES: Olawale O.E. Ajibola, N/A, Ph.D, University of Lagos: Grant/Research Support