Demographic characteristics prediction using deep learning analysis of kidney imaging

Evaluation of kidney health is often done by measuring abnormal levels of creatinine and albumin; however, they only measure secondary artifacts of larger changes in kidney health. Having only an indirect measure leads to the need for more rounds of testing, such as nuclear medicine scans. Kidney function, in particular, is often related to patient demographics and comorbidities.1 Age-related changes in kidney structure result in decreased estimated glomerular filtration rates. Differences in kidney size exist between men and women, leading to variations in creatinine levels. Given the correlation between underlying kidney function and patient demographics, it may be possible to leverage demographic variables as proxy variables for underlying kidney function. Giyocha et al. demonstrated identification of patient demographics from imaging using AI was effective, surpassing the ability of clinicians.2 Previous works have established changes in organ appearance, such as white matter changes seen in brain MRI, are correlated with patient age.3 Furthermore, deviations by the model when estimating patient age could be related to the onset of disease, providing an effective mechanism for monitoring disease progression. In this work, we seek to develop models to estimate/classify both patient age and gender using image patches containing kidneys. The kidney parenchyma degrades due to the patient’s age, with certain diseases exaggerating the degradation. We aim to demonstrate the ability to reliably identify patient demographic features while using model errors to identify unhealthy individuals.