Disputation: Large data and machine learning in analysis, diagnostics, and clinical decision making: applications in the treatment of burn injury
- Plats: Rudbecklaboratoriet Fåhraeussalen, ingång C5
- Doktorand: Jian Fransén, doktorand vid institutionen för kirurgiska vetenskaper, plastikkirurgi
- Kontaktperson: Fredrik Huss
Jian Fransén försvarar sin avhandling "Large data and machine learning in analysis, diagnostics, and clinical decision making: applications in the treatment of burn injury”. Disputationen kommer att hållas på engelska.
Burn injury is a common trauma globally. Large burns require fluid resuscitation, infection control, and specialized intensive care. The size of the burn and infections caused by resistant microbes are correlated to mortality, and accurate mortality predictions are important. Errors are common when diagnosing burn depth, but early diagnosis is necessary to make correct surgical decisions. Machine learning (ML) is a set of mathematical algorithms with self-learning capabilities, which might make them suitable for medical applications.
This thesis explores systematic large data analysis and ML algorithms for clinical applications in burn treatment by examining antibiotic resistance, improving mortality predictions, and automating diagnosis of burn depth.
Paper I aims to find relevant trends and correlations on clinical outcomes such as mortality, microbial distribution, and antibiotic resistance from pooled data from a burn center. Data from 1570 patients and 15,006 microbiology cultures were systematically analyzed. Our results show a sustained low risk of harmful microbes, resistance, and a suggested low mortality rate.
Paper II used clinical biomarkers from burn patients to train ML algorithms to predict mortality and compare it with Baux scores. When applying five types of ML algorithms, it showed no significant difference in mortality prediction compared with Baux scores.
Paper III examines convolutional neural network (CNN) algorithms for two purposes. One to segment a burn wound and the other to classify whole wound images for surgery or conservative treatment. A total of 1105 diverse images were collected from patients at admission to burn centers in Sweden and South Africa. The algorithm was adequate for segmenting burn wounds and could be improved when categorizing images for surgery or conservative treatment.
Paper IV further assesses CNN to automatically segment and diagnose a diverse set of early burn images for deep or superficial burn injury. A total of 1004 images were included. The algorithm proved adequate in segmenting superficial injuries but not deep injuries and performed similarly between darker and lighter skin patients.
Future studies might incorporate infection variables in ML mortality predictions and larger sample sizes. Regarding automated burn image diagnosis, including multiple non-image variables might improve usability.
Länk till avhandlingen i DiVA: https://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-513553