A carregar...
Publicação
Multimodal Machine Learning for Early Detection of Clinical Deterioration in Chronic Heart Failure
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 8.11 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A Insuficiência Cardíaca Crónica (ICC) é uma das principais causas de hospitalização e
mortalidade, sendo frequentes os episódios de deterioração clínica de curto prazo. Este
projeto investiga a utilização de técnicas de Mineração de Dados, Machine Learning e
Processamento de Linguagem Natural para apoiar a predição de deterioração clínica de curto
prazo em doentes com ICC estabelecida, integrados num programa de monitorização remota.
Uma Revisão Sistemática da Literatura, apoiada pela estrutura PRISMA 2020 (Preferred
Reporting Items for Systematic Reviews and Meta-Analyses), estabeleceu o contexto clínico
e metodológico. Seguindo a metodologia CRISP-DM (Cross-Industry Standard Process for
Data Mining), foi desenvolvido um enquadramento multimodal utilizando dados de 181
doentes com ICC. Medições fisiológicas estruturadas e dados de alertas foram combinados
com notas clínicas não estruturadas. O texto foi limpo e processado através da extração de
conceitos clínicos baseada em regras, representações TF-IDF baseadas em uni-, bi- e
trigramas, e embeddings contextuais gerados com o BioBERTpt-clin, um modelo linguístico
clínico português baseado em BERT. Os sintomas, comorbilidades, tratamentos e outros
indicadores derivados do texto foram agregados ao nível do doente e integrados com
características fisiológicas numa matriz de risco clinicamente fundamentada.
Foram desenvolvidos modelos de Regressão Logística regularizados para prever
instabilidade clínica grave numa janela de 90 dias: um modelo base estruturado, baseado em
indicadores fisiológicos, e um modelo integrado que incorporou conceitos derivados do
texto. Um classificador Random Forest foi avaliado como referência comparativa não linear.
A validação cruzada estratificada a cinco partições produziu valores de ROC-AUC de 0,855
para o modelo estruturado e 0,857 para o modelo integrado.
Os indicadores de instabilidade fisiológica foram os preditores mais fortes, enquanto as
características derivadas do texto acrescentaram contexto clínico. Apesar do enquadramento
retrospetivo de implementação única, os resultados demonstram a viabilidade da
estratificação de risco multimodal no apoio à decisão baseado em telemonitorização na
gestão da ICC.
Chronic Heart Failure (CHF) is a leading cause of hospitalization and mortality, with patients frequently experiencing short-term clinical deterioration. This project investigates the use of Data Mining, Machine Learning, and Natural Language Processing to support short-term deterioration prediction in patients with established CHF enrolled in a remote patient monitoring programme. A Systematic Literature Review (SLR), supported by the PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework, established the clinical and methodological context. Following the CRISP-DM methodology (Cross Industry Standard Process for Data Mining), a multimodal framework was developed using data from 181 CHF patients. Structured physiological measurements and warning data were combined with unstructured clinical notes. Text was cleaned and processed using rule-based clinical concept extraction, TF-IDF representations based on uni-, bi-, and tri-grams, and contextual embeddings generated using BioBERTpt-clin, a Portuguese clinical BERT-based language model. The extracted symptoms, comorbidities, treatments and other text-derived indicators were aggregated at patient level and integrated with physiological features in a clinically grounded risk matrix. Regularized Logistic Regression models were developed to predict severe clinical instability within a 90-day window: a structured physiological baseline and an integrated model incorporating text-derived concepts. A Random Forest classifier was evaluated as a non linear benchmark. Five-fold stratified cross-validation produced ROC-AUC values of 0.855 for the structured model and 0.857 for the integrated model. Physiological instability indicators were the strongest predictors, while text-derived features added relevant clinical context. Despite the retrospective single-deployment setting, the findings demonstrate the feasibility of multimodal risk stratification to support telemonitoring-based decision support in CHF management.
Chronic Heart Failure (CHF) is a leading cause of hospitalization and mortality, with patients frequently experiencing short-term clinical deterioration. This project investigates the use of Data Mining, Machine Learning, and Natural Language Processing to support short-term deterioration prediction in patients with established CHF enrolled in a remote patient monitoring programme. A Systematic Literature Review (SLR), supported by the PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework, established the clinical and methodological context. Following the CRISP-DM methodology (Cross Industry Standard Process for Data Mining), a multimodal framework was developed using data from 181 CHF patients. Structured physiological measurements and warning data were combined with unstructured clinical notes. Text was cleaned and processed using rule-based clinical concept extraction, TF-IDF representations based on uni-, bi-, and tri-grams, and contextual embeddings generated using BioBERTpt-clin, a Portuguese clinical BERT-based language model. The extracted symptoms, comorbidities, treatments and other text-derived indicators were aggregated at patient level and integrated with physiological features in a clinically grounded risk matrix. Regularized Logistic Regression models were developed to predict severe clinical instability within a 90-day window: a structured physiological baseline and an integrated model incorporating text-derived concepts. A Random Forest classifier was evaluated as a non linear benchmark. Five-fold stratified cross-validation produced ROC-AUC values of 0.855 for the structured model and 0.857 for the integrated model. Physiological instability indicators were the strongest predictors, while text-derived features added relevant clinical context. Despite the retrospective single-deployment setting, the findings demonstrate the feasibility of multimodal risk stratification to support telemonitoring-based decision support in CHF management.
Descrição
Palavras-chave
Insuficiência cardíaca crónica Machine learning Processamento de linguagem natural Predição de risco Mineração de dados Telemonitorização