Clinical decision support models for oropharyngeal cancer treatment: design and evaluation of a multi-stage knowledge abstraction and formalization process

• Published in: International journal for computer assisted radiology and surgery Vol. 17; no. 9; pp. 1643 - 1650
• Main Authors: Gaebel, Jan, Mehlhorn, Stefanie, Oeser, Alexander, Dietz, Andreas, Neumuth, Thomas, Stoehr, Matthaeus
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2022; Springer Nature B.V
• Subjects: Bayesian analysis; Cancer; Complexity; Computer Imaging; Computer Science; Decision support systems; Expert systems; Health Informatics; Imaging; Iterative methods; Knowledge engineering; Machine learning; Mathematical models; Medicine; Medicine & Public Health; Original; Original Article; Parameters; Pattern Recognition and Graphics; Radiology; Surgery; Vision; Chemotherapy; Patient-specific modeling; Clinical decision support systems; Cancer of the oropharynx
• ISSN: 1861-6429; 1861-6410; 1861-6429
• DOI: 10.1007/s11548-022-02675-3

Purpose Treatment decisions in oncology are demanding and affect survival, general health, and quality of life. Expert systems can handle the complexity of the oncological field. We propose the application of a hybrid modeling approach for decision support models consisting of expert-based implementation of a decision model structure and machine-learning (ML) based parameter generation. We demonstrate our approach for the treatment of oropharyngeal cancer. Methods We created a clinical decision model based on Bayesian Networks and iteratively optimized its characteristics using structured knowledge engineering approaches. We combined manual adaptation of individual concepts with automatic learning of parameters and causalities. Using data from 94 patient records, we targeted the needed objectivity and clinical significance. Results In three iteration steps, we assessed the model with cross-validations. The initial aggregated accuracy of 0.529 could be increased to 0.883 in the final version. The predictive rates of the target nodes range from 0.557 to 0.960. Conclusion Combining different methodological approaches requires balancing the complexity of the clinical subject matter with the amount of information available in the dataset for ML application. Our method showed promising results because flaws of one approach can be overcome by the other approach. However, technical integrability as well as clinical acceptance must always be ensured.