Disability and Imaging AI: Performance of Chest X-Ray (CXR)-based DL Model in patients with and without disabilities

Purpose: Most deep learning (DL) models are applied to patients without disabilities or with sparse representation of disabled patients. We assessed the performance of a commercial DL model for detecting CXR findings in patients with and without chronic musculoskeletal disabilities affecting the upper body (UBD+).

Methods/Materials: Our retrospective, IRB-approved study included 1,005 patients (mean age 57±18years) with 1,005 posteroanterior CXRs (507 patients with UBD [+UBD] and 498 without UBD [UBD-]) from 5 community and 2 quaternary care hospitals and MIMIC CXR database. Patients with UBD+ were identified from the Epic Warehouse Database (EDW) SQL server. Radiology reports were obtained from the radiology reports search engine (Nuance mPower). Next, we extracted the presence or absence of the following DL assessable findings in the radiology reports: atelectasis, pneumonia, pulmonary nodule, pneumothorax, pleural effusion, and cardiomegaly. All CXRs were processed with the DL model. Data were analyzed with receiver operating characteristics (ROC) analysis using area under the curve (AUC) separately for each finding. De Long analysis was used to compare performance between both groups.

Results: The distribution of findings between UBD- (AI/reported findings ratio: 108/152 atelectasis, 255/195 pneumonia, 196/115 pulmonary nodule, 85/81 pneumothorax, 144/149 pleural effusion, and 76/34 cardiomegaly) and UBD groups (113/108 atelectasis, 215/96 pneumonia, 109/15 pulmonary nodules, 3/1 pneumothorax, 83/80 pleural effusion, and 90/27 cardiomegaly) varied significantly. DL had significantly higher AUCs for identifying any CXR finding in UBD- [0.865 (95% CI 0.820-0.910)] than in UBD+ [0.732 (95% CI 0.688-0.776)], p=0.001. For individual findings, a significant difference was found in DL performance for lung nodules [UBD+ 0.6 (95% CI 0.439-0.751), UBD- 0.773 (95% CI 0.725-0.822), p=0.016] and pleural effusion [UBD+ 0.833 (95% CI 0.773-0.893), UBD- 0.916 (95% CI 0.883-0.950), p=0.006].

Conclusions: There are substantial and statistically significant differences in CXR-DL model performance in patients with and without upper body disabilities. We will extend the study to multiple DL models to confirm such bias.