Mining medicine prescriptions to enhance predictive management of COVID-19 patients in the UAE

Abstract

Abstract Data mining is becoming more popular in the healthcare industry to enhance decision-making and patient outcomes. In this study, we examined medication patterns in over 10,000 COVID-19 patients spanning over three years from the UAE. Prescriptions digitization enabled their utility as data evidence for analytical and predictive modeling tools including machine learning (ML). In this work, we attempt to utilize digitized free-text prescriptions associated with 10k + COVID-19-positive hospitalized cases, to first extract a three-layered hierarchy of administered medicines and then use them as data features to understand their administration patterns, reveal the impacts on and associations with patients' treatment to improve the performance of predicting the key treatment outcomes. We determined higher frequencies of certain medications during different stages of the pandemic and discovered correlations between medication co-administration habits and patient outcomes such as ICU admission, ventilator usage, prolonged hospital stays, and mortality. Next, we train ML models for forecasting targeted variables and found that MEWS and Age are associated with increased risk for ICU admission, ventilator use, lengthy hospital stays, and mortality. A singleton micro-predictor of the selected feature targets in connection with the likelihood of death reveals that VENTILATOR_USE and IS_SEPTIC feature elevates the likelihood of death to 60%. Remarkably, the use of ANTISPASMODIC and ANTIFUNGAL have high mortality rates and support (PX 0.151,0.055, P DEATH/X 0.379, 0.641, respectively). Underscoring the importance of data-driven approaches to inform clinical decision-making. The application of clustering and co-administration patterns generated from graph theory may offer cutting-edge pandemic control techniques in preparation for the next pandemic.