Machine Learning techniques applied to identifying clinical factors regarding automated medical prognosis

• Autor(a) principal: Barros, Pedro Henrique Ferracini de
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-19012023-093014/
• topic: Árvores de decisão; Attention mechanisms; Clinical prediction explicability; Clinical trajectory prediction; Clusterização hierárquica; Construção de cohorts de pacientes; Decision trees; Electronic Health Records; Encoder- Decoder; Encoder-Decoder; Explicabilidade de predições clínicas; Fenotipação de pacientes; Hierarchical clustering; Mecanismos de atenção; Patient cohort construction; Patient phenotyping; Predição de trajetórias clínicas; Prontuários Médicos Eletrônicos

Machine Learning Algorithms have shown promising results in several areas of knowledge, including preventive medicine. While deep learning techniques have been shown to be effective for automated medical prognosis, they lack more transparency and interpretability. On the other hand, data clustering techniques and decision trees are promising for the identification of risk factors, characteristics in common, and trends among patients according to their respective clinical stories, described by their Electronic Health Records (EHRs). In this sense, this MSc thesis aimed to elaborate a machine learning framework composed of an Attentive EncoderDecoder neural network to predict patients next admissions diagnoses, a hierarchical clustering algorithm to phenotype these predictions, and finally, a decision tree aiming the phenotypes explicability; each step of our framework produced particular results: the Attentive EncoderDecoder obtained state-of-the-art results over the datasets MIMIC-III and MIMIC-IV-ED; the clustering algorithm produced consistent results of related diagnoses of one same phenotype and, also, neighboring phenotypes demonstrated to have similar diagnoses; and finally, the decision tree provided a visualization of the decision rules between diagnoses of phenotypes and demonstrated the irrelevance of patients demographic data in comparison to their diagnoses when identifying a phenotype. We summarize our contributions as follows: (i) Achievement of state-of-the-art results with a versatile model based on an Attentive Encoder-Decoder, named by us AttentionHCare (BARROS; RODRIGUES, 2022), (ii) Provision of a decision support tool to specialists as an explainable model, and (iii) Ability to identify patterns (e.g. risk factors, common diagnoses, bias, etc.) in patients with similar clinical trajectories by using the explainable model.