Sex bias exists in diagnosing pulmonary arterial hypertension via machine learning

Abstract

Objective: Pulmonary arterial hypertension (PAH) is more common in females, yet males tend to have a worse prognosis, indicating potential sex-based differences. Will such differences affect the PAH diagnosis accuracy by machine learning (ML) on cardiac magnetic resonance imaging (CMRI)?

Methods: A retrospective cohort of 220 consecutive subjects with PAH (105 F; 45 M) or with no pulmonary hypertension (51 F; 19 M) were included in the study. An ML pipeline [1] was implemented using the baseline CMRI Short Axis and 4 Chamber scans for PAH diagnosis prediction utilizing (a) the full cohort and (b) only female subjects. Fisher's exact test with Bonferroni correction was used to compare the proportionate accuracy between the sexes on 150 PAH subjects.

Results: No significant difference was found between the sexes using right heart catheterisation PAH diagnosis (p=0.75). In the full cohort (a), ML-predicted diagnosis demonstrated a bias towards correctly predicting PAH in males on both the Short Axis (p=0.01) and 4 Chamber (p<0.01). The Short Axis ML-predicted diagnosis correctly classified 132 PAH subjects (88/105 F; 44/45 M). The 4 Chamber ML-predicted diagnosis accurately classified 128 PAH subjects (83/105 F; 45/45 M). In the sex-stratification ML pipeline (b), accuracy was improved on both the Short Axis (89/105, i.e. +1) and 4 Chamber (93/105, i.e. +10).

Conclusions: This preliminary study found significant disparity in PAH prediction accuracy by ML between the sexes. This suggests that sex bias exists in CMRI-based PAH diagnosis via ML and sex-stratification could be benefitial in ML-based diagnosis.

[1] Swift, A, et al. Eur. Heart J. Cardiovasc. Imaging 2020; jeaa001