To identify possible risk factors for stroke-like events in patients with the m.3243A>G pathogenic variant.

Mitochondrial encephalomyopathy, lactic acidosis, and stroke-like episodes (MELAS) syndrome is a devastating mitochondrial disorder, frequently caused by the mitochondrial DNA (mtDNA) m.3243A>G variant. Not all carriers of the m.3243A>G variant develop stroke-like events and there are currently no clear predictors of disease severity.

Clinical and laboratory data were analyzed from a prospective natural history study enrolling healthy controls, carriers of the m.3243A>G variant with MELAS, and symptomatic carriers without MELAS between 2015 and 2018. Standard statistical analyses were conducted to identify differences among groups. Principal Component Analysis (PCA) was employed to uncover key patterns and features across the groups. Random Forest classifiers were trained on a supervised predictive task to discriminate patients with MELAS from controls; carriers without MELAS were excluded from training.

The cohort included 37 controls, 86 carriers of the m.3243A>G variant without stroke-like episodes at enrollment, and 22 carriers with MELAS. Among carriers without MELAS, 6 developed stroke-like episodes during the study (converters). Comparison of converters and non-converters revealed significant differences in heteroplasmy (urinary sediment), 6-minute-walk test (distance) and Karnofsky score. PCA using clinical and routine laboratory features revealed clear separation between patients with MELAS and controls, providing evidence for their inclusion in predictive models. On inference on symptomatic carriers, our machine-learning model was able to forecast stroke occurrence in 4 of the 6 converters. Key predictors included elevated m.3243A>G heteroplasmy, low Karnofsky score, and severity of muscle and CNS involvement.

Defining progression of disease in carriers of the m.3243A>G variant and identifying risk-predictors of stroke is crucial for patients’ counseling and clinical trials planning. Our model suggested that clinical and conventional laboratory data may be sufficient to predict disease severity thus allowing a broad applicability in research and clinical practice.