The development of efficient screening tools for endocrine disorders may have clinical impacts, both in terms of improved prognoses of individual patients through disease detection at an earlier stage and the cost-effective allocation of public health resources by focusing on individuals with a high risk of disease and avoiding unnecessary testing in low-risk groups. Researchers have sought to determine whether ML algorithms are able to provide a better way of screening for various endocrine diseases. Artzi et al. [4] provided an excellent example of applying the principles of ML to find useful screening tools for gestational diabetes based on a sizeable electronic health record (EHR) database. EHR data of 588,622 pregnancies from 368,351 women collected at the nationwide level in Israel between 2010 and 2017 were used to train an ML model to predict the risk of gestational diabetes. Among 2,355 candidate features, the researchers developed a simple model consisting of only nine self-reportable questions (without previous laboratory results in some cases) based on a gradient boosting model, which showed fair discriminatory performance (area under the receiver operating characteristic curve, 0.80 vs. 0.68 for the conventional glucose challenge test at 24 to 28 weeks of gestation) even at an earlier time point relative to the initiation of pregnancy. Medical image data have the potential to provide features suitable for the opportunistic screening of endocrine disorders. Valentinitsch et al. [12] trained an ML model to identify individuals with prevalent vertebral fractures based on non-fractured vertebral regions in computed tomography scans taken for various purposes. By combining global and local density and texture parameters, the ML model outperformed volumetric bone mineral density (BMD) alone in discriminating the presence of vertebral fractures, suggesting the potential of a semi-automated pipeline for the opportunistic screening of individuals with a high risk of fracture. Kong et al. [13] developed an ML model to detect facial features from photos of patients with acromegaly, and their model may have the potential to help the detection of acromegaly at an earlier stage.

Tackling the gray area of diagnostic uncertainty with new modalities has always been an important task for clinicians. Asymptomatic hyperparathyroidism can be challenging to identify without a high index of suspicion because it involves subtle biochemical changes and its phenotype overlaps with those of primary osteoporosis and other rare mineral disorders, including familial hypocalciuric hypercalcemia [27]. Somnay et al. [25] trained an ML model to identify patients with primary hyperparathyroidism among patients who underwent neck surgery, including thyroidectomy or parathyroidectomy, although relatively low performance was shown for mild disease. Several studies have shown that ML could support the decision process of whether to perform an invasive biopsy on a thyroid nodule based on ultrasonography, with good classification performance similar to that of radiology experts; therefore, ML classifications might potentially provide guidance to operators during data acquisition and measurement [1428]. A well-validated, accurate, non-invasive ML model may have the potential to replace standard invasive diagnostic modalities for certain diseases. For instance, the global burden of nonalcoholic fatty liver disease (NAFLD) is rapidly growing, but invasive liver biopsy remains the gold standard for diagnosing NAFLD and nonalcoholic steatohepatitis. Perakakis et al. [15] developed a support vector machine-based model to classify NAFLD based on features obtained from the lipidomic, glycomic, and liver fatty acid analysis of serum samples. For the presence of liver fibrosis, a parsimonious exploratory model with 10 lipid species showed high accuracy (up to 98%), suggesting the possibility of a targeted lipidomic approach as an alternative non-invasive diagnostic tool, although the model needs to be further validated in other ethnicities and individuals with a milder spectrum of liver diseases [29].

Although unsupervised learning has been utilized less often than supervised learning for diagnosis and screening, it may be helpful to find novel clusters and associations within a given dataset. Kruse et al. [16] applied unsupervised hierarchical agglomerative clustering to find groups with high and low risks of fracture in women from a national Danish patient database based on BMD, medication reimbursement, anthropometric characteristics, and comorbidities. Among the nine clusters that were identified, four clusters classified as corresponding to a high risk of fracture showed heterogeneous compliance to antiresorptive treatments, even with a similar distribution of BMD. The age of 60 years was the earliest time point that allowed a clear discrimination between high and average fracture risk. Altogether, that study provided novel insights regarding characteristics related to compliance with bone medications and the optimal age to recommend dual-energy X-ray absorptiometry screening.

Accurately predicting clinical outcomes enables an individualized approach to treatment strategy and monitoring. The Weight, Age, hyperTension, Creatinine, High-density lipoprotein cholesterol, Diabetes control, and Myocardial infarction (WATCH-DM) score was developed to predict heart failure risk among patients with type 2 diabetes using ML algorithms based on the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial dataset, and showed good predictive performance with an external validation set (the Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial [ALLHAT]) [17]. Su et al. [18] found that a simple model including only age and BMD selected by classification and regression tree analysis performed similarly to Fracture Risk Assessment Tool (FRAX) categories as a reference tool for predicting incident hip fracture in a large cohort of community-dwelling older men.

ML principles can be used to find specific subgroups with a heterogeneous response to treatment. Basu et al. [19] re-analyzed the ACCORD trial data to find subgroups with different treatment effects in response to intensive glucose control compared to standard therapy. Although intensive glucose control was associated with increased mortality in the ACCORD trial published in 2008, their post hoc analysis identified that a subgroup of patients experienced a survival benefit from intensive treatment, and the proportion of patients in the subgroup that made the main contribution to increased mortality in the trial was relatively small. This study provides an example of the utility of ML in dissecting treatment responses, with the potential for a more tailored approach both for interpreting results from previous trials and for applying therapeutic strategies according to individual status. For the prediction of treatment response in patients with acromegaly, anthropometric and biochemical data with imaging features were combined in an ML model that achieved better prognostication than the reference prediction tool [20]. Good ML models may have the potential to provide guidance for dose adjustment, particularly in patients with chronic conditions requiring the indefinite replacement of certain hormones, as in patients who receive thyroid hormone replacement after total thyroidectomy or in type 1 diabetes patients who receive insulin replacement. Zaborek et al. [21] built a supervised ML model to guide levothyroxine dose adjustment, which showed a fair improvement of predictive accuracy compared to the current standard of weight-based dosing. A reinforcement training algorithm has been applied to guide the optimal dosing of long-acting insulin in patients with type 1 diabetes [5]. Although the results are preliminary, these studies illustrate the ongoing efforts made by endocrinology researchers to improve patient care by achieving better predictions of the disease course and response to treatment.