Abstract Rationale Pulmonary tuberculosis continues to be an important public health problem in India and many other lower middle income (LMIC) countries. Post-tuberculosis lung disease (PTLD) is characterized by the presence of respiratory symptoms attributable at least in part to previous tubercular infection. The presence of respiratory symptoms in patients with a prior history of tuberculosis can be suggestive of PTLD, but recurrence and other complications such as chronic pulmonary aspergillosis (CPA) need to be considered since they warrant specific treatment. Computed tomography (CT) scans help differentiate between these conditions better (such as between healed and active tuberculosis) but require trained manpower for their interpretation. We aimed to develop and validate an artificial intelligence (AI) model to diagnose post-tuberculosis lung disease (PTLD) based on chest CT scans and to differentiate it from recurrence of tuberculosis and chronic pulmonary aspergillosis (CPA), and to assess the accuracy of this model when compared to a clinico-radio-microbiological gold standard. Methods We enrolled patients with a history of pulmonary tuberculosis and current symptoms of fever, cough or loss of weight. Microbiological and serological tests were performed on serum and respiratory specimens and CT scans of the chest were performed and interpreted by an experienced radiologist. A diagnosis was established based on a combined clinical, radiological and microbiological composite reference standard (CRS). A convoluted neural network (CNN) was trained based on these diagnoses and CT scans. Results A total of 137 patients were enrolled; 96 scans were used to train the model (PTLD - 44, CPA - 36, recurrent PTB - 16) and external validation was performed on the remaining 41 scans. Of the external validation cohort, 32 patients had PTLD, 7 had CPA and 2 had recurrent pulmonary tuberculosis. The model had a sensitivity of 90.63 % with a positive predictive value (NPV) of 76.3%. However, its specificity and negative predictive value were disappointing, possibly owing to the skewed distribution of patients with alternative diagnoses (CPA and recurrent pulmonary tuberculosis) in the external validation cohort. Conclusions Our model was able to achieve acceptable sensitivity at diagnosing PTLD, but its specificity was limited by the skewed distribution of patients’ diagnoses in the external validation cohort. Fig. 1: Confusion matrix showing the true diagnoses on the y axis and our model’s prediction on the y axis; labels: 0- CPA, 1- Recurrent TB, 2 - PTLD CONCLUSIONS: Our model was able to achieve acceptable sensitivity at diagnosing PTLD, but its specificity was limited by the skewed distribution of patients’ diagnoses in the external validation cohort. This abstract is funded by: None
Palani et al. (Fri,) studied this question.