Predicting coronavirus disease 2019 severity using explainable artificial intelligence techniques

Predicting coronavirus disease 2019 severity using explainable artificial intelligence techniques

Abstract Predictive models for determining coronavirus disease 2019 (COVID-19) severity have been established; however, the complexity of the interactions among factors limits the use of conventional statistical methods. This study aimed to establish a simple and accurate predictive model for COVID-...

Full description

Saved in:
Bibliographic Details
Main Authors: Takuya Ozawa, Shotaro Chubachi, Ho Namkoong, Shota Nemoto, Ryo Ikegami, Takanori Asakura, Hiromu Tanaka, Ho Lee, Takahiro Fukushima, Shuhei Azekawa, Shiro Otake, Kensuke Nakagawara, Mayuko Watase, Katsunori Masaki, Hirofumi Kamata, Norihiro Harada, Tetsuya Ueda, Soichiro Ueda, Takashi Ishiguro, Ken Arimura, Fukuki Saito, Takashi Yoshiyama, Yasushi Nakano, Yoshikazu Muto, Yusuke Suzuki, Ryuya Edahiro, Koji Murakami, Yasunori Sato, Yukinori Okada, Ryuji Koike, Makoto Ishii, Naoki Hasegawa, Yuko Kitagawa, Katsushi Tokunaga, Akinori Kimura, Satoru Miyano, Seishi Ogawa, Takanori Kanai, Koichi Fukunaga, Seiya Imoto
Format: Article
Language:English
Published: Nature Portfolio 2025-03-01
Series:Scientific Reports
Subjects:
COVID-19
Machine learning
Artificial intelligence
Online Access:https://doi.org/10.1038/s41598-025-85733-5
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract Predictive models for determining coronavirus disease 2019 (COVID-19) severity have been established; however, the complexity of the interactions among factors limits the use of conventional statistical methods. This study aimed to establish a simple and accurate predictive model for COVID-19 severity using an explainable machine learning approach. A total of 3,301 patients ≥ 18 years diagnosed with COVID-19 between February 2020 and October 2022 were included. The discovery cohort comprised patients whose disease onset fell before October 1, 2020 (N = 1,023), and the validation cohort comprised the remaining patients (N = 2,278). Pointwise linear and logistic regression models were used to extract 41 features. Reinforcement learning was used to generate a simple model with high predictive accuracy. The primary evaluation was the area under the receiver operating characteristic curve (AUC). The predictive model achieved an AUC of ≥ 0.905 using four features: serum albumin levels, lactate dehydrogenase levels, age, and neutrophil count. The highest AUC value was 0.906 (sensitivity, 0.842; specificity, 0.811) in the discovery cohort and 0.861 (sensitivity, 0.804; specificity, 0.675) in the validation cohort. Simple and well-structured predictive models were established, which may aid in patient management and the selection of therapeutic interventions.
ISSN:2045-2322

Similar Items