Abstract:
Mortality prediction models play a pivotal role in evaluating patients’survival prospects within defined time frames.In the context of IntensiveCare Units (ICUs), the accurate prediction of patient mortality has garneredlong-standing interest.The ICU setting is critical for patients requiringintensive care to maximize their chances of survival, particularly thosewith moderate survival probabilities.Early mortality prediction holdsthe potential to identify vulnerable patients and mitigate life-threateningrisks. In this thesis, we focus on predicting the mortality of ICU patientsusing machine learning (ML) approaches,employing the largest publiclyavailable ICU data, namely Medical Information Mart for Intensive Care(MIMIC)-III database [1].Predicting early ICU mortality poses challengesdue to the complex, high-dimensional, irregular, and imbalanced nature ofICU data.Moreover, the time-series records for ICU patients add furthercomplexity. While risk stratification models have evolved into more intricateand precise models based on the expanding availability of medical data,data integration from disparate sources remains a challenge.Furthermore,healthrecordscontainaconsiderablylargeamountofmissingdata,meaningthat not all patients have all types of features/measurements – making theapplication of ML models difficult. Therefore, selecting suitable models forpredicting the mortality of specific ICU cohorts is important yet challenging.We developed prediction models applicable across diverse patient profiles,aiding timely decisions and improving patient outcomes.We addressedvarious challenging scenarios that may frequently arise in ICU data andused appropriate techniques to circumvent these challenges.We proposedthree techniques to impute the missing values and evaluated their efficacy inmortality prediction. We evaluated a compendium of ML techniques, pairedwitheffectivetechniquesforhandlingmissingdataandclassimbalance,topredictthemortalityofICUpatients.Weaimedatpredictingtherisk of the death of a patient 12 hours in advance.Our results indicatethat, despite various challenges, reasonably accurate prediction is possiblegiven the data is well curated and imputed with appropriate techniques.Overall, this thesis contributes to understanding the challenges associatedwith mortality prediction of diverse ICU patient cohorts, developing newapproaches to handle the challenging issues that frequently arise in thefield, and showcasing the power of ML approaches in assisting medicalpractitionersinfastandaccuratedecision-making.
