AI APPLICATIONS IN CLINICAL PRACTICE OF ENDOCRINOLOGY : AN OVERVIEW
Susan Raghavan , MD , FAAP
ARTIFICIAL INTELLIGENCE
Artificial intelligence ( AI ) is the theory and development of computational systems and algorithms that are designed to accomplish tasks that normally require human intelligence . The term is somewhat nebulous , also encompassing the idea of computerized intelligence and conceptualization of human experience . AI is often designed to mimic human neurological structures in a computational form . The most common example of this is neural network , within which each “ node ” can be thought of as representative of an individual neuron , with weighted connections analogous to axonal connections between neurons .
Within medicine , a subset of AI called machine learning ( ML ) has been applied most broadly to multiple specialties . Machine learning involves the development of algorithms designed to accomplish a singular task most commonly from a large data set . Machine learning algorithms are trained via a multitude of different methods , but one commonly utilized method is called supervised learning . These algorithms are trained via reinforcement learning in which the algorithm is essentially trained by providing a data set and a known output or answer desired from the algorithm . A calculation called back propagation is used to alter the weights and connections of the algorithm to obtain said output from the given data set . Repeating this process over thousands of data sets is called supervised learning . A few examples of this include algorithms for predicted survival following a multitude of different cancers , to algorithms intended to develop strategies for insulin pump closed-loop systems which rely on large data sets of prior blood glucose measurements to titrate and administer insulin based on an individual ’ s historical response . 1 , 2 , 3
AI / ML has revolutionized the world of medicine . The very first AI-based technologies approved by the FDA were in the field of radiology ( software analyzing cardiovascular images from MR : 2016 ), ophthalmology ( ophthalmic application to screen for diabetic retinopathy in April 2018 which was the first system used in a clinical setting ), dermatology ( diagnostic software for lesions suspicious for cancer : 2019 ) to name a few . Also in 2018 , the Guardian Connect System from Medtronic was approved for blood sugar predictions from trans-dermal sensors . 4
Ever-growing contributions of machine intelligence to endocrine-imaging for tumors of pancreas , adrenal , pituitary and thyroid have great potential to enable diagnostics to preclude or limit invasive biopsy . ML techniques in tumor imaging have demonstrated the potential to differentiate hormone secreting from non-hormone secreting pituitary adenomas and demarcate tumor from normal tissue in about 30 minutes , enabling intra-operative tumor delineation improving patient outcomes . ML algorithms have demonstrated higher sensitivity , specificity and positive and negative predictive values in analyses of multiple disease states .
AI / ML technologies in endocrinology have been increasingly evaluated in :
1 ) Screening and diagnosis
AI in gestational diabetes : Using large datasets from pregnant women in Israel between 2010 and 2017 , ML tools were able to not only predict pregnancies likely to be complicated by gestational diabetes , but were also able to identify women at risk for gestational diabetes pre-pregnancy with discriminatory accuracy . 5
AI / ML in vertebral fractures : Utilizing characteristics of both fractured bone and the non-fractured vertebral regions in CT spine outperformed volumetric bone density ( BMD ). Facial recognition studies were able to detect early changes of acromegaly using photographs of patients with and without the disease . 3
AI / ML in growth disorders in children : Algorithms for height monitoring can now be integrated into electronic health records , which can increase the diagnostic yield and identify individual children who may have growth failure .
Assessment of the cause of growth failure in a child frequently includes cranial imaging . A recent development is the routine inclusion of T2-DRIVE into sellar MRI protocols . In light of recent safety concerns regarding gadolinium contrast agents , the computer-aided technique is considered a valid alternative for pituitary imaging without gadolinium in patients with pituitary hormone deficiencies . This method has been more accurate in the diagnosis of pituitary gland abnormalities since it has been shown to provide better contrast than gadolinium agents . However , it should always be noted normative data of pituitary dimensions is required for interpretation of any technique and good clinical expertise remains ( continued on page 12 ) APRIL 2022 11