Titulo Estágio
Confidence and personalization of machine learning models : application to cardiovascular risk assessment
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
Laboratório de Computação Adaptativa do CISUC
Enquadramento
Machine Learning (ML) and Artificial Intelligence (AI) methods have achieved notable performance in many fields, and its research impact in healthcare is unquestionable. Nevertheless, the deployment of intelligent models in clinical practice is somewhat still limited, namely in the risk assessment of cardiovascular patients. Some of the major issues, acknowledged as barriers to successful real-world applications, include lack of personalization, explainability and confidence. These aspects are definitely decisive not only for patient safety, but also to promote the trustworthiness of professionals.
The explainability of clinical models is critical to obtain confidence and adherence of professionals, while facilitating the integration of available different sources of information. In particular, the generation of a set of explanations, by combining data-driven methods with clinical evidence, to obtain transparent models will allow circumventing the black-box limitations of common ML models.
Additionally, to increase the trustfulness of a model, besides the understanding of its behaviour, clinicians should also have confidence on the predicted outcomes. Therefore, the development of reliability/confidence measures, able to complement the predicted outcome of a ML model with an estimation for its confidence, is of fundamental importance.
Finally, current risk models are typically designed using the “one fits all” principle, lacking a truly personalization. Additionally, they fail the adequate integration of clinical evidence and knowledge from a holistic clinical practice. Personalized models, with the ability to combine existent risk models with new information (new risk factors) and patients’ specific information to match the most appropriate characteristics of a patient, thus avoiding the principle “one fits all”, would represent a significant added value.
Objetivo
Several risk assessment models are available to be applied by the physicians in the daily clinical practice. This project aims to improve the management of myocardial infarction (MI) patients, through the development of predictive cardiovascular risk assessment models in order to predict the likelihood of death and re-hospitalization of patients with an episode of acute MI. In this context, this proposal will research on computational intelligence techniques, explainable artificial intelligence methods and data fusion approaches for addressing personalization, explainability and reliability issues. In particular, a strategy developed inside the group will be improved.
The main scientific hypothesis to validate is the possibility to holistically complement current risk assessment models used in clinical practice with the most recent advances in artificial intelligence, to come up with better and trustworthy risk models
By tackling these challenges, it will contribute to increase the acceptability and trust of professionals regarding such models, thus contributing not only to effectively assist clinicians in their decisions, but also to incorporate these models in clinical workflows.
Plano de Trabalhos - Semestre 1
State of the art on concepts related with the topic
- Explainable machine learning techniques
- Confidence measurements for machine learning models
- Personalization techniques
- Cardiovascular risk assessment scores
Preliminary results
- Confidence measurements for machine learning models
Plano de Trabalhos - Semestre 2
- Research and develop explainable machine learning techniques applied to cardiovascular risk assessment
- Research and develop personalization techniques applied to risk assessment
- Implement confidence measurements for machine learning models applied to risk assessment
- Thesis writing
For validation purposes a retrospective study based on two available datasets will be carried out: the largest MI Portuguese dataset (N=16000), provided by SPC (to be confirmed); the data set provided by CHUC (N=2000).
Condições
The work will take place at the Laboratories of the Adaptive Computing Group at CISUC, at DEI.
There is the possibility of a scholarship, depending on the approval of a project
Observações
Jorge Henriques, jh@dei.uc.pt, DEI - CISUC
Simão Paredes, sparedes@isec.pt, ISEC- CISUC
Dr. José Pedro Sousa, zpedro_14@hotmail.com, CHUC – Centro Hospitalar e Universitário de Coimbra.
Orientador
Jorge Henriques
jh@dei.uc.pt 📩