Uri Nahum (1), Julie Refardt (2), Irina Chifu (3), Wiebke Kristin Fenske (4,5), Martin Fassnacht (3,6), Gabor Szinnai (1), Mirjam Christ-Crain (2) and Marc Pfister (1)
(1) University of Basel Children’s Hospital (UKBB), Basel, Switzerland; (2) Departments of Endocrinology, Diabetology and Metabolism, University Hospital Basel, Basel, Switzerland, (3) Division of Endocrinology and Diabetes, Department of Internal Medicine I, University Hospital, University of Wuerzburg, Wuerzburg Germany, (4) Integrated Research and Treatment Center for Adiposity Diseases, Leipzig University Medical Center, Leipzig, Germany, (5) Leipzig University Medical Center, IFB Adiposity Diseases, Leipzig, Germany, (6) Central Laboratory, University Hospital Wuerzburg, Wuerzburg, Germany
Objectives:
Differentiation between central Diabetes Insipidus (cDI) and Primary Polydipsia (PP) remains challenging in clinical practice. Although the hypertonic saline infusion test led to a high diagnostic accuracy [1], it is a laborious test requiring close monitoring of plasma sodium levels and it causes considerable discomfort for patients. As such, we leverage Machine Learning (ML) methods to simplify and facilitate the diagnosis of cDI.
Methods:
In this work, we analyzed clinical data from 59 patients with cDI and 81 patients with PP from a prospective multi-center study evaluating the hypertonic saline test as a new test approach to diagnose cDI. Our primary outcome was the diagnostic accuracy of the ML based algorithm (such as artificial neural networks, logistic regression and k-nearest neighbours) in differentiating cDI from PP patients. The data set used included 56 clinical, biochemical, and radiological covariates. We identified a set of five covariates that were crucial for differentiating cDI from PP patients utilizing ML methods. We developed ML based algorithms on the data and after optimizing them on a training-data set, we validated them using an in advance separated unseen test-data set.
Results:
Applying statistical methods, we identify five parameters as input for the basic ML based algorithm: urine osmolality, plasma sodium and glucose, known transphenoidal surgery or anterior pituitary deficiencies. Testing the ML algorithm on the unseen test-data set resulted in a high Area Under the Curve (AUC) score of 0.87. A further improvement of the ML based algorithm was reached by choosing a suitable additional MRI characteristic (stalk enlargement) or by adding the results of the hypertonic saline infusion test (AUC score of 0.93 and 0.98, respectively).
Conclusions:
The developed ML based algorithm has the potential to simplify and facilitate differentiation between cDI and PP patients with a high accuracy with clinical and laboratory data only, thereby possibly avoiding MRI or other cumbersome clinical tests in the future. Although the developed ML based cDI algorithm cannot replace physicians and their experience when performing diagnosis, such algorithm may serve as complimentary decision support tools facilitating identification of pediatric and adult patients that should undergo more complex diagnostic testing. Our covariate selecting process and the ML based approach could be developed for other diseases, especially in the field of adult and pediatric endocrinology.
References:
[1] Fenske W, Refardt J, Chifu I, Schnyder I, Winzeler B, Drummond J, Ribeiro-Oliveira A Jr, Drescher T, Bilz S, Vogt DR, Malzahn U, Kroiss M, Christ E, Henzen C, Fischli S, Tönjes A, Mueller B, Schopohl J, Flitsch J, Brabant G, Fassnacht M, Christ-Crain M. A Copeptin-Based Approach in the Diagnosis of Diabetes Insipidus. N Engl J Med. 2018 Aug 2;379(5):428-439.
Reference: PAGE 30 (2022) Abstr 10117 [www.page-meeting.org/?abstract=10117]
Poster: Methodology – AI/Machine Learning