Pieroni, M (2023) Machine Learning Modelling of Critical Care Patients in the Intensive Care Units. Doctoral thesis, Liverpool John Moores University.

Preview

Text 2023PieroniPhD.pdf - Published Version Available under License Creative Commons Attribution Non-commercial. Download (5MB) | Preview

Abstract

The ICU is a fast-paced data-rich environment which treats the most critically ill patients. On average, over 15 % of patients admitted to the ICU amount in mortality. Therefore, machine learning (ML) is paramount to aiding the optimisation and inference of insight in critical care. In addition, the early and accurate evaluation of the severity at the time of admission is significant for physicians. Such evaluations make patient management more effective as they are more likely to predict whose physical conditions may worsen. Moreover, ML techniques could potentially enhance patients' experience in the clinical setting by providing medical alerts and insight into future events occurring during hospitalisation. The need for interpretable models is crucial in the ICU and clinical setting, as it is vital to explain a decision that leads to any course of action related to an individual patient.

This thesis primarily focuses on mortality, length of stay forecasting, and AF classification in critical care. We cover multiple outcomes and modelling methods whilst using multiple cohorts throughout the research. However, the analysis conducted throughout the thesis aims to create interpretable models for each modelling objective. In Chapter 3, we investigate three publicly available critical care databases containing multiple modalities of data and a wide range of parameters. We describe the processes and contemplations which must be considered to create actionable data for analysis in the ICU. Furthermore, we compared the three data sources using traditional statistical and ML methods and compared predictive performance. Based on 24 hours of sequential data, we achieved AUC performances of 79.5% for ICU mortality prediction and a prediction error of approximately 1.3 hours for ICU LOS.

In Chapter 4, we investigate a sepsis cohort and conduct three sub-studies. Firstly, we investigated sepsis subtypes and compared biomarkers using traditional modelling methods. Next, we compare our approach to commonly and routinely used scoring systems in the ICU, such as APACHE IV and SOFA. Our tailored approach achieved superior performance with pulmonary and abdominal sepsis (AUC 0.74 and 0.71respectivly), displaying distinct individualities amongst the different sepsis groups. Next, we further expand our analysis by comparing ML methods and inference approaches to our baseline model and ICU acuity scores. We further investigate extending analysis to other outcomes of interest (In-hospital/ICU mortality, In-hospital/ICU LOS) to gain a more holistic view of the sepsis derivatives. This research shows that nonlinear models such as RF and GBM commonly outperformICU scoring, methods such as APACHE IV and SOFA and linear methods such as logistic/linear regression. Lastly, we extend our analysis in a multi-task learning framework for model optimisation and improved predictive performance. Our results showed superior performance with pulmonary, abdominal and renal/UTI sepsis (AUC 0.76, 0.77 and 0.73, respectively). Lastly, Chapter 5 investigates the classification of atrial fibrillation (AF) in long-lead ECG waveforms in sepsis patients. We developed a deep neural network to classify AF ECGs from Non-AF ECG cases in conjunction with refining a method to gain insight from the neural network model. We achieved a predictive performance of 0.99 and 0.89 regarding the test and external validation data. The inference from the model was achieved through the use of saliency maps, dimensionality reduction methods and clustering, utilising the automatic features learned by the developed model. We developed visualisations to help support the inference behind the classification of each ECG prediction.

Overall, the research displays a wide range of novelties and unique approaches to solving various outcomes of interest in the ICU. In addition, this research demonstrates the implication of ML applicability in the ICU environment by providing insight and inference to diverse tasks regardless of the level of complexity. With further development, the frameworks and approaches outlined in this thesis have the potential to be used in clinical practice as decision-support tools in the ICU, allowing the automated alert and detection of patient classification, amongst others. The results generated in this thesis resulted in journal publications and medical understanding gained from insight available in the developed ML frameworks.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Machine Learning; AI; ICU; Sepsis; Cardiovascular; Atrial Fibrillation; Critical Care; Intensive Care
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > R Medicine (General)
Divisions:	Computer Science & Mathematics
SWORD Depositor:	A Symplectic
Date Deposited:	19 Jun 2023 10:11
Last Modified:	19 Jun 2023 10:12
DOI or ID number:	10.24377/LJMU.t.00019587
Supervisors:	Oiler, I, Ortega-Martorell, S, Lip, G and Welters, I
URI:	https://researchonline.ljmu.ac.uk/id/eprint/19587

View Item