Filza Noor Asari1*, Muhammad Faiz Mohd Hisham1, Eida Nurhadzira Muhammad1, Noor Aliza Lodz1, Chong Zhuo Lin1
1. Institute for Public Health, National Institutes of Health, Ministry of Health, Setia Alam, 40170, Shah Alam, Selangor, Malaysia
*Corresponding author: Filza Noor Asari.
ABSTRACT
INTRODUCTION: Artificial Intelligence (AI) tools, such as Qure.AI and Putralytica, offer rapid and automated chest X-ray (CXR) interpretations, which are valuable in TB-endemic areas. However, limited data exist on how patient-level factors influence these tools’ classification behaviour.
OBJECTIVE: To examine how sociodemographic and clinical characteristics influence the PTB classification by two AI-based tools, Qure.AI and Putralytica.
MATERIALS & METHODS: This retrospective study analysed 1308 CXRs classified as Pulmonary Tuberculosis (PTB), Normal, and Abnormal non-PTB. CXRs, patient demographics, and clinical history were retrieved from the Institute for Clinical Research. The dependent variable was each AI tool’s classification of CXR as PTB versus non-PTB. Multiple logistic regression identified predictors of PTB classification, including age, gender, ethnicity, comorbidities, TB-related symptoms, and smoking status. Adjusted odds ratios (aORs) were computed to determine the associated factors, independent of ground-truth diagnosis.
RESULTS: Qure.AI’s performance in PTB classification was significantly reduced among patients aged ≥60 years (aOR=0.64; 95% CI: 0.48-0.86, p=0.002), TB-related symptoms (aOR=0.68; 95% CI: 0.51-0.93, p=0.014), and comorbidity: patients with ≥1 comorbidity (aOR=0.18; 95% CI: 0.12-0.26, p<0.001). In contrast, smokers had higher PTB classification odds than non-smokers (aOR=2.03; 95% CI: 1.39-2.96, p<0.001). Similar predictors were significant for Putralytica, with reduced performance in older age ≥60 years (aOR=0.45; 95% CI: 0.35-0.58, p<0.001), and comorbidity (aOR=0.64; 95% CI: 0.50-0.82, p<0.001). Male patients had higher odds of PTB classification than females (aOR=1.36; 95% CI: 1.06-1.76, p=0.017). Smoking status was associated with increased odds (aOR=1.39; 95% CI: 1.03-1.87, p=0.031).
CONCLUSION: Provided the study cohort is representative and free from significant selection bias, these findings demonstrate that age, comorbidity, smoking, and gender influence the classification behaviour of selected AI tools. This analysis offers a crucial evaluation of model fairness, which is distinct from diagnostic accuracy. Findings underscore the need for algorithm auditing and subgroup performance monitoring to ensure equitable deployment of AI tools in TB screening.
KEYWORDS: Pulmonary Tuberculosis, Artificial Intelligence, Sociodemographic, Clinical Determinants


