Artificial Intelligence (AI) and machine learning (ML) in risk prediction of hospital acquired pressure injuries (HAPIs) among oncology inpatients.

Abstract

e18095 Background: Utilizing AI and ML is an emerging method to improve risk identification, characterization and stratification for clinical outcomes such as HAPIs. The Jvion Cognitive Clinical Success Machine (CCSM) utilizes the Eigen Sphere technique to factor in clinical, socioeconomic, and behavioral covariates at the individual patient level to maximize accuracy of risk prediction and provide insights on prevention of HAPIs. Methods: A retrospective analysis was performed utilizing claims and EHR data on 63,476 inpatient admissions between June 2016 and June 2018 at M D Anderson Cancer Center, a 660-bed oncology facility (Table). All risk assessment indicators in the data were removed prior to analysis by the CCSM to compute unbiased risk probabilities for HAPIs. The performance of the CCSM risk prediction for all new stage 2 and above HAPIs ( > = Stage 2 HAPIs) was compared with the Braden scale using AUC. Results: The Jvion CCSM had an AUC of 0.84 compared to the AUC of 0.72 for the Braden scale in the prediction of > = Stage 2 HAPIs. Conclusions: The AUC value indicates that the Jvion CCSM has better predictive accuracy than the Braden scale. It also has better discrimination in risk identification. Thus, Jvion CCSM can be a valuable tool in risk screening for HAPIs which can lead to early preventative interventions. [Table: see text]

Machine Learning Approaches for Hospital Acquired Pressure Injuries: A Retrospective Study of Electronic Medical Records

“…A recent study applied six diverse machine learning algorithms to a cohort of 7,717 Intensive Care Unit (ICU) patients and reported a C-statistic of 0.83 ( 11 ), while another study reported a C-statistic of 0.84 for a general hospital population of 8,286 observations using logistic regression with under-sampling of the control patients during model fitting ( 12 ). Other studies include: (1) application of Bayesian Network approaches to the aforementioned cohort of 7,717 ICU patients, achieving a similar C-statistic of 0.83 as before, while improving sensitivity and adding model interpretation through modeling of related risk factors (e.g., medications, diagnoses, and Braden scale factors) ( 13 ), (2) a random forest model which leveraged predictors curated from clinical input and previous literature to predict stage 1 HAPI and above with a C-statistic of 0.79 ( 14 ), (3) another logistic regression which leveraged ICU-specific features to obtain a recall of 0.74 ( 15 ), and (4) other modeling approaches built off of Electronic Medical Records (EMR) and claims data, an online AI platform and another logistic regression model (after comparison between six machine learning methods), obtaining a C-statistic of 0.84 and recall of 0.67, respectively ( 16 , 17 ). In the Supplementary Material , we have included a table which summarizes these studies for the purpose of comparison to the current study's findings ( Supplementary Table 1 ).…”

Investigating the potential for machine learning prediction of patient outcomes: a retrospective study of hospital acquired pressure injuries

“…A 2015 article applied six diverse machine learning algorithms to a cohort of 7,717 ICU patients and reported a C-statistic of 0.83 [11], while another study reported a C-statistic of 0.84 for a general hospital population of 8,286 observations using logistic regression with under-sampling of the negative cases during model fitting [12]. Other studies have applied Bayesian Network approaches to Braden subscales [13], random forest [14] and models built off of Electronic Medical Records (EMR) and claims data [15,16].…”