DEVELOPMENT OF A MACHINE LEARNING-BASED MODEL FOR SEPSIS PREDICTION IN INTENSIVE CARE UNITS

Sepsis is a life-threatening disorder that may easily escalate to the deterioration of organs and death unless early detected and addressed. Early detection of sepsis among patients in Intensive Care Unit (ICU) is thus, important in enhancing outcomes. It is a paper that introduces a machine learning model based on Long Short-Term Memory (LSTM) networks to predict the onset of sepsis based on temporal trends on patient vital signs, laboratory tests, and demographic information. The eICU Collaborative Research Database data were pre-processed, resampled, and missing value imputed, normalised, and sequence generated (as an LSTM input). The LSTM model was trained in 15 epochs and tested on another test set. The highest scores were an AUROC of 0.91, a precision of 0.78, a recall of 0.82, and F1-score of 0.80, which predicts sepsis 3.5 hours on average before clinical diagnosis. The confusion matrix showed that the model was quite useful in recognising non-sepsis patients and achieving high accuracy in most cases of sepsis but there were also false negativity. These findings also indicate that LSTM-based networks are able to model temporal relationships in ICU patient records in order to predict early signs of sepsis, which has the potential of providing an opportunity to intervene in the process of clinical intervention. This paper demonstrates the opportunity to unify predictive models based on machine learning with critical care units, as it is a solution that can help clinicians make decisions grounded in data and lead to better patient outcomes. Future researchers are recommended to improve the sensitivity and add some clinical characteristics and to test the model on various ICU populations to guarantee its wide applicability.