USING A NOVEL ARTIFICIAL INTELLIGENCE (AI) BASED APPROACH TO ANALYZE THE CONSEQUENTIAL RISE IN NON-COVID-19 MORTALITY DURING THE COVID-19 PANDEMIC

ABSTRACT

SESSION TITLE Use of Machine Learning and Artificial Intelligence SESSION TYPE Original Investigations PRESENTED ON 10/16/22 1030 am - 1130 am PURPOSE: The COVID-19 pandemic has significantly impacted the US healthcare system. Between March 1, 2020, and January 2, 2021, a 22.9% increase in all-cause mortality was reported [1]. We used Artificial Intelligence (AI) for data analysis to have a prototype national average by matching various characteristics. This is a novel approach known as Digital Twinning Method (DTM). We intend to compare non-COVID mortality between 2020 and 2019 using this DTM approach. METHODS: Data was collected by a contracted vendor that provided analysis utilizing an AI framework. Mortality rates were calculated at four points of care categorized as 1) In-patient mortality, 2) 30-day on-admission, 3) 30-day on discharge, and 4) 90-day on-admission. Baseline risk predictions were generated using DTM for matching patient demographics such as age, gender, race, Medicare status, and community-dwelling status. Hence, each person was compared to a "twin” with the same risk of hospitalization, death, acute myocardial infarction, or stroke. RESULTS: Our institution had a higher actual non-COVID mortality in 2020 compared to the actual mortality in 2019 across all four points of care studied. The highest increase was noticed in the 90-day on-admission category (9.7% in 2019 vs 12.6% in 2020) followed by 30-day on-admission (5.0% in 2019, to 6.6% in 2020), 30-day on-discharge (4.2% in 2019, to 5.7% in 2020), and in-patient mortality (1.8% in 2019, to 2.6% in 2020). However, when compared to twinned patients at other hospitals, our institution had a lower non-COVID mortality rate across all categories in 2019 and 2020. We utilized the Sign Test to evaluate our repeated-paired-measures for the above four points of care categories during two different conditions, i.e., under a normal healthcare situation (2019) and in the pandemic year (2020). Our two-tailed p-value was 0.0455 with statistical significance at p < 0.05, with M1-M2 (M=measure) difference of -0.8 (in-patient mortality), -1.6 (30 day on-admission), -1.5 (30 day on-discharge), and -2.9 (90 day on-admission) for the four categories. Our z-score was +2 under the formula z = (X - pn) / √npq, signifying positive deviation from the mean. Our study was limited by the unavailable data of patients who may have had COVID but were undiagnosed. CONCLUSIONS: AI is a novel method to obtain reliable data. Based on our results, we conclude that the non-COVID mortality rate at our institution increased during the pandemic. Further studies are needed to specify the underlying causes attributable to the increased mortality. CLINICAL IMPLICATIONS By leveraging Artificial intelligence in healthcare to analyze big datasets and perform complex analyses, it may be of clinical importance to utilize AI-generated risk prediction models to accurately identify variables that can be controlled in future pandemics to decrease mortality while increasing overall efficiency of the healthcare system. DISCLOSURES No relevant relationships by Muhammad Mohsin Abid No relevant relationships by Sana Jogezai No relevant relationships by Iqbal Ratnani No relevant relationships by Muhammad Hassan Virk No relevant relationships by Anza Zahid