A commentary published this past July in The Journal of Clinical Endocrinology & Metabolism asked, “Can machine learning transform the diagnostic approach to congenital disorders of adrenal steroidogenesis?”
Jani Liimatta, MD, PhD, of the Institute of Clinical Medicine at the University of Eastern Finland in Kuopio, wrote the paper opining on another study published in JCEM by Tosun et al titled, “Machine learning algorithms to accelerate etiological diagnosis of congenital disorders of adrenal steroidogenesis.”
Tosun and her co-authors point out that congenital disorders of adrenal steroidogenesis (CDAS) comprise a heterogeneous group of inherited conditions characterized by dysregulated biosynthesis of cortisol, aldosterone, and mineralocorticoids in the adrenal cortex. These disorders can lead to substantial short- and long-term adverse health consequences, they write.
That team developed and validated a machine-learning decision tree model, using a development cohort of 1027 participants (325 genetically confirmed CDAS patients representing eight subtypes/702 controls) for model construction – the model achieved a mean overall accuracy of 97.1%. “Machine learning–assisted steroid profiling provides an accurate and highly interpretable diagnostic approach for CDAS, with potential for integration into pediatric endocrine diagnostics and decision-support systems,” Tosun and her co-authors concluded.
Liimatta starts his commentary by noting how crucial it is for patients to receive a timely and accurate diagnosis of CDAS. “Congenital disorders of adrenal steroidogenesis represent a diverse group of inherited conditions in which timely and precise etiological diagnosis is essential for optimal clinical management,” he writes.
“The integration of high-resolution steroid profiling with an interpretable [machine-learning] framework represents a significant and timely advance in the diagnostic approach to disorders of adrenal steroidogenesis. The study by [Tosun et al.] provides a well-executed example of how these complementary methodologies can be combined to enhance the interpretation of complex biochemical data.”
Liimatta goes on to explain that clinicians often rely on sequential hormonal testing, dynamic stimulation studies, and confirmatory genetic analyses, all of which may be time-consuming and dependent on specialized expertise. “These challenges are particularly consequential in pediatric settings, where delays in diagnosis may lead to life-threatening complications or inappropriate therapeutic decisions,” he writes.
Liimatta praises the work of Tosun and her co-authors, calling it a thoughtful and important contribution to the field. He then reflects on several key insights from their work: that Tosun et al. incorporated multiple steroids into an integrated analytical framework, which “captures broader patterns of pathway disruption, more closely reflecting the underlying physiology of adrenal steroidogenesis,” the use machine learning as a structured analytical layer since it offers the offers the potential to support and standardize the interpretation of complex hormonal data, and how the study by Tosun et al. emphasized interpretability.
Liimatta does point to a couple of limitations with their work – small sample sizes, the use of imputation for missing data. But a lot of times, such is the case when studying rare disorders.
In his conclusion, Liimatta writes that he sees machine learning should be viewed as a tool to augment and support clinical expertise, rather than something to replace it. “In summary,” he writes, “the integration of high-resolution steroid profiling with an interpretable [machine-learning] framework represents a significant and timely advance in the diagnostic approach to disorders of adrenal steroidogenesis. The study by [Tosun et al.] provides a well-executed example of how these complementary methodologies can be combined to enhance the interpretation of complex biochemical data.”
