Predicting Psychiatric Diseases Using AutoAI: A Performance Analysis Based on Health Insurance Billing Data
Zotero / D&S Group / Top-Level Items 2022-02-23
Type
Conference Paper
Author
Markus Bertl
Author
Peeter Ross
Author
Dirk Draheim
Editor
Christine Strauss
Editor
Gabriele Kotsis
Editor
A. Min Tjoa
Editor
Ismail Khalil
Series
Lecture Notes in Computer Science
Place
Cham
Publisher
Springer International Publishing
Pages
104-111
ISBN
978-3-030-86472-9
Date
2021
DOI
10.1007/978-3-030-86472-9_9
Library Catalog
Springer Link
Language
en
Abstract
Digital transformation enables a vast growth of health data. Because of that, scholars and professionals considered AI to enhance quality of care significantly. Machine learning (ML) algorithms for improvement have been studied extensively, but automatic artificial intelligence (autoAI/autoML) has been widely neglected. AutoAI aims to automate the complete AI lifecycle to save data scientists from doing low-level coding tasks. Additionally, autoAI has the potential to democratize AI by empowering non-IT users to build AI algorithms. In this paper, we analyze the suitability of autoAI for mental health screening to detect psychiatric diseases. A sooner diagnosis can lead to cost savings for healthcare systems and decrease patients’ suffering. We evaluate AutoAI using the open-source machine learning library auto-sklearn, as well as the commercial Watson Studio’s AutoAI platform to predict depression, post-traumatic stress disorder, and psychiatric disorders in general. We use health insurance billing data from 83,986 patients with a total of 687,697 ICD-10 coded diseases. The results of our research are as follows: (i) on average, an accuracy of 0.6 (F11_{\!{1}}–score 0.58) with a precision of 0.61 and recall of 0.56 was achieved using auto-sklearn. (ii) The evaluation metrics for Watson Studio’s autoAI were 0.59 accuracy, 0.57 F11_{\!{1}}–score, a precision of 0.6, and a recall of 0.55. We conclude that the prediction quality of autoAI in psychiatry still lacks behind traditional ML approaches by about 24% and is therefore not ready for production use yet.
Proceedings Title
Database and Expert Systems Applications
Short Title
Predicting Psychiatric Diseases Using AutoAI