JMIR Publications

ABSTRACT

Background:

Today, machine learning enabled clinical decision support research and applications generate actionable insights by utilizing large amounts of ICU patient data and the generated clinical decision support suggestions could be applied to various ICU clinical scenarios. Machine learning, sometimes called data-driven method, develops multiple statistical analysis models using computational technologies. It allows computer systems to learn from patient data and discover unknown clinical situations. There are considerable previous studies in ICU scenario to conduct prediction or guide the treatment of critically ill patients.

Objective:

We sought to achieve an overall review of research and application of machine learning enabled clinical decision support studies in intensive care units (ICU). Our review study may help ICU clinicians, researchers, developers and policy makers better understanding advantages and limitations of machine learning supported diagnose, outcome prediction, risk events identification or treatment recommendations ICU point of care.

Methods:

In total of 643 articles published on PubMed between Jan. 1980 and Oct. 2020 were collected. According to our selection criteria, 97 studies that focused on machine learning enabled clinical decision support studies in intensive care units were selected and reviewed on the following aspects: research topics, study cohorts, machine learning models, analysis variables, and evaluation methods.

Results:

All the selected studies were categorized into 4 topics with 13(13.4%) monitoring, detection and diagnosis, 32(33.0%) Early identification of clinical events, 46(47.6%) outcome prediction and prognosis assessment and 6(6.2%) treatment decisions related articles. 82(84.5%) studies were developed with data from adult patients, 9(9.3%) with pediatric patients and 6(6.2%) was for neonates. 65(67.0%) studies used data from a single center and 32(33.0%) studies were developed from a multi-center dataset. 88(90.7%) studies used supervised learning, only 3 (3.1%) studies used unsupervised learning whereas 6 (6.2%) studies used reinforcement learning. The ranking of the clinical variable categories distribution in these studies are: demographics (74), lab values (59), vital signs (55), scores (48), ventilation parameters (43), comorbidities (27), medications (18), outcome (14), fluid balance (13), non-medicine therapy (10), symptoms (7) and medical history (4). The most frequently adopted evaluation methods for clinical data modeling studies are AUROC (61), sensitivity (51), specificity (41), accuracy (29) and positive predictive value (23).

Conclusions:

We demonstrated a comprehensive review of the machine learning enabled clinical decision support studies in intensive care units. Research and application status of intelligent machine learning models, methods and results were reviewed. Besides, limitations, challenges and future directions also discussed in our study.