Dense neural networks in knee osteoarthritis classification: a study on accuracy and fairness

• Published in: Neural computing & applications Vol. 35; no. 1; pp. 21 - 33
• Main Authors: Moustakidis, Serafeim, Papandrianos, Nikolaos I., Christodolou, Eirini, Papageorgiou, Elpiniki, Tsaopoulos, Dimitrios
• Format: Journal Article
• Language: English
• Published: London Springer London 01.01.2023; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Arthritis; Artificial Intelligence; Body weight; Classification; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Diagnosis; Feature extraction; Image Processing and Computer Vision; Intelligence; Knee; Machine learning; Medical research; Neural networks; Osteoarthritis; Probability and Statistics in Computer Science; Risk analysis; S.I. : Information; Special Issue on Information; Subgroups; Systems and Applications; Data classification; Knee osteoarthritis; Fairness; Machine learning; Dense neural networks
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-020-05459-5

Dense neural networks (DNNs) are a powerful class of learning algorithms that uses multiple layers to progressively extract higher level features from raw input. Either deep or shallow, their outstanding capabilities made a very significant impact on improving the diagnostic potential in multiple applications including medical data classification. In this research work, DNN and Machine Learning (ML) models are explored to address the diagnosis problem of knee osteoarthritis classification which is a common complex problem in older adults. Knee OA diagnosis is a highly complex problem being related to a large number of medical risk factors including advanced age, gender, hormonal status, body weight or size, family history of disease, etc. The main research objective of this study is to apply DNN in knee osteoarthritis classification and validate it for the first time with respect to both accuracy and fairness. To accomplish this, a hybrid criterion including accuracies, confusion matrix and two fairness metrics (demographic parity (DP) and balanced equalized odds (BEO)) were employed to validate the performance of the proposed methodology. Different subgroups of control participants from self-reported clinical data were considered to prove the performance of the proposed methodology. The best performing DNN method is compared with some popular and well-known machine learning techniques for classification with respect to accuracy and fairness. The results of the conducted experimental analysis show the efficacy of the proposed DNN approach improving the classification accuracy (up to 79.6%) and fairness (BEO: ~ 92% and DP: 98.5%) in the OA case study.