Predicting treatment outcome in cardiac TAVI procedures through machine learning
Marco Mamprin defended his PhD thesis at the Department of Electrical Engineering on October 29th.
Aortic valve stenosis, a leading cause of mortality among the elderly, is increasingly treated with Transcatheter Aortic Valve Implantation (TAVI). With his PhD research Marco Mamprin aims to improve patient selection for TAVI by leveraging machine learning to predict mortality more accurately, addressing the limitations of current risk models. Starting with a retrospective analysis, key mortality predictors were identified and validated using cross-validation and federated learning across two hospitals. Additionally, ECG analysis identified leads associated with post-TAVI pacemaker needs, enhancing pre-procedure risk assessment. Practical prototypes were developed to support clinical integration, and a dedicated data-sharing protocol ensures compliance with privacy standards. Ultimately, these innovations aim to improve TAVI outcomes by refining high-risk patient identification and supporting informed clinical decisions.
Transcatheter Aortic Valve Implantation (TAVI) has emerged as a routine treatment option; however, careful patient selection remains critical for its success. Candidates for TAVI frequently present with multiple comorbidities, necessitating improved methods to identify those at higher risk of post-procedure complications.
Current risk models exhibit limited accuracy in predicting TAVI outcomes. Several models, which are non-specific to TAVI, have been developed and are currently used to estimate the risk of procedural or 30-day mortality. This research shifts focus to one-year mortality, which is considered more relevant for TAVI patients. The thesis aims to leverage advanced machine learning techniques to enhance patient selection processes and minimize associated mortality.
Valuable feature insights
Marco Mamprin began his research with a retrospective analysis of clinical data from 270 patients at Catharina Hospital Eindhoven. By utilizing state-of-the-art model-explanation techniques based on decision trees, significant features influencing one-year mortality were identified, leading to the development of an initial predictive model. The model interpretation provides valuable feature insights that enable medical professionals to supervise and assess the model鈥檚 decision-making process.
In collaboration with the Amsterdam UMC, an extended study established a dedicated exchange protocol to overcome data exchange issues due to individual center policies. Extended cross-validation were conducted, validating multiple models using various techniques on two populations, involving 631 and 1,300 patients per center. Additionally, federated learning techniques were employed, demonstrating the potential for improved outcomes by training models iteratively using data from different centers.
Prototypes well-received
The researcher also explored electrocardiogram (ECG) data to predict the need for permanent pacemaker installation post-TAVI. Analyzing 2,197 ECGs from 631 patients, key predictive leads were identified at three stages: pre-procedure, during-procedure, and post-procedure.
Practical challenges related to implementing a predictive model in clinical practice are addressed through the development of two prototypes: a web interface where patient data can be input to query a predictive model, and an API that queries the model and provides prediction-related information. Both prototypes were well-received by medical professionals, highlighting the importance of validating machine learning models in clinical settings.
Data-sharing protocol
In conclusion, predicting TAVI outcomes remains a significant challenge, particularly concerning one-year survival rates. Through this research, the validation process was extended to other populations from different hospitals to assess the generalizability of the proposed solution. Data sharing limitations were overcome by developing a validation protocol that exchanges models and data processing pipeline, instead of data. Finally, the use of ECG data was exploited to identify patients at higher risk of conduction disorders after the procedure, further reducing complications.
The research was performed under a EU project called project.
The project received an award called: Winner of ITEA Award of Excellence Exploitation in 2021.
Title of PhD thesis: . Supervisors: Prof. Peter de With, and Dr. Sveta Zinger.