Feature extraction and heart failure detection in ex vivo perfusion systems
(2026)Department of Automatic Control
- Abstract
- This thesis investigates the detection of acute heart failure (AHF) in an exvivo heart perfusion system using hemodynamic signals. Early detection of AHF is important for improving monitoring and intervention. Experimental data was pre-processed and transformed into a beat-to-beat representation, from which physiological features were extracted. Events were labeled based on manual reductions in pump flow supported by left ventricular pressure (LVP) analysis.
Three detection models were developed using a flag-based framework combining threshold-based features: an all-feature model, a reduced model, and a weighted model emphasizing the most informative signals. Performance was evaluated using F2 score, confusion matrices, and latency.
All... (More) - This thesis investigates the detection of acute heart failure (AHF) in an exvivo heart perfusion system using hemodynamic signals. Early detection of AHF is important for improving monitoring and intervention. Experimental data was pre-processed and transformed into a beat-to-beat representation, from which physiological features were extracted. Events were labeled based on manual reductions in pump flow supported by left ventricular pressure (LVP) analysis.
Three detection models were developed using a flag-based framework combining threshold-based features: an all-feature model, a reduced model, and a weighted model emphasizing the most informative signals. Performance was evaluated using F2 score, confusion matrices, and latency.
All models were able to detect approximately 80-90% of AHF events. The weighted model achieved the most consistent performance, with higher F2 scores and better generalization to test data, while the reduced model showed a larger performance drop on the test data. The reduced and weighted models produced approximately one false positive per true positive, whereas the allfeature model exhibited a higher ratio of about 1.4 false positives per true positive. All models were able to detect events prior to intervention.
The study is limited by the small dataset size and potential bias from using LVP for both labeling and feature extraction. The results should therefore be interpreted as exploratory, but demonstrate the potential of combining hemodynamic features for early AHF detection. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9225614
- author
- Nilsson, Hanna
- supervisor
- organization
- year
- 2026
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6307
- other publication id
- 0280-5316
- language
- English
- id
- 9225614
- date added to LUP
- 2026-05-11 09:31:39
- date last changed
- 2026-05-11 09:31:39
@misc{9225614,
abstract = {{This thesis investigates the detection of acute heart failure (AHF) in an exvivo heart perfusion system using hemodynamic signals. Early detection of AHF is important for improving monitoring and intervention. Experimental data was pre-processed and transformed into a beat-to-beat representation, from which physiological features were extracted. Events were labeled based on manual reductions in pump flow supported by left ventricular pressure (LVP) analysis.
Three detection models were developed using a flag-based framework combining threshold-based features: an all-feature model, a reduced model, and a weighted model emphasizing the most informative signals. Performance was evaluated using F2 score, confusion matrices, and latency.
All models were able to detect approximately 80-90% of AHF events. The weighted model achieved the most consistent performance, with higher F2 scores and better generalization to test data, while the reduced model showed a larger performance drop on the test data. The reduced and weighted models produced approximately one false positive per true positive, whereas the allfeature model exhibited a higher ratio of about 1.4 false positives per true positive. All models were able to detect events prior to intervention.
The study is limited by the small dataset size and potential bias from using LVP for both labeling and feature extraction. The results should therefore be interpreted as exploratory, but demonstrate the potential of combining hemodynamic features for early AHF detection.}},
author = {{Nilsson, Hanna}},
language = {{eng}},
note = {{Student Paper}},
title = {{Feature extraction and heart failure detection in ex vivo perfusion systems}},
year = {{2026}},
}