1. Title of the paper: 'A Novel Comparison of Artificial Intelligence Methods for Diagnosing Knee Osteoarthritis' 2. Name: Mr Luca Parisi Mailing address: Via Celle 24 B, Villar Dora (TO), 10040, Italy. E-mail address: lpar962@aucklanduni.ac.nz. Phone number: +64 223917807. Name: Mr Paul Robert Biggs Mailing address: Cardiff School of Engineering, The Queen's Buildings, The Parade, Cardiff, South Glamorgan CF24 3AA, United Kingdom. E-mail address: BiggsP1@cardiff.ac.uk . Phone number: +44 2920874716. Name: Dr Gemma Marie Whatling Mailing address: Room W/2.34, Cardiff School of Engineering, The Queen's Buildings, The Parade, Cardiff, South Glamorgan CF24 3AA, United Kingdom. E-mail address: WhatlingGM@cardiff.ac.uk. Phone number: +44 2920876348. Name: Professor Catherine Avril Holt Mailing address: Room W/2.45, Cardiff School of Engineering, The Queen's Buildings, The Parade, Cardiff, South Glamorgan CF24 3AA, United Kingdom. E-mail address: Holt@cardiff.ac.uk. Phone number: +44 2920874533. 3. Mr Luca Parisi 4. Abstract of the paper: Dr Jones successfully developed a technique at Cardiff University that facilitates the classification between pathological and non pathological knee function using Dempster-Shafter theory (DST), demonstrating its effectiveness in comparison to Artificial Neural Networks (ANN) and Linear Discriminate Analysis (LDA) in the classification of osteoarthritic gait. To the best of our knowledge, this study proposes, for the first time, to compare the performance and suitability of the (i) Cardiff Classifier, (ii) ANN and (iii) SVM in discriminating between healthy and osteoarthitic knee function. Thirty-eight patients with late stage knee osteoarthritis (OA) and thirty-eight healthy volunteers were instructed to complete walking trials at speeds they deemed to be normal. Spatio-temporal gait data (stride length, cadence, body mass index) and principal components of knee kinematic and ground reaction force waveforms were averaged for five gait cycles. Eighteen selected biomechanical and clinical variables formed the input matrix for the three classifiers. After randomising the data, the physiological knee function of healthy volunteers and pathological function due to knee OA were used as desired outputs for the following learning classifiers systems: - a four-dimensional unsupervised Self-Organising Map (SOM) designed via MatLab (Mathworks inc., USA); - a supervised Multi-layer Perceptron (MLP) neural network with one hidden layer having 20 hidden neurons designed through MatLab (Mathworks inc., USA); - Lagrangian Support Vector Machines (LSVMs) Machine Learning (ML)-based classification method and - the Dempster-Shafter theory (DST)-based classifier (‘Cardiff classifier’). To prevent over-fitting/overtraining, that is to ensure that the learning classifier systems were truly able to discriminate between healthy and pathological (with OA) patterns of the gait for diagnosis, a ten k-fold (70-15-15) cross-validation algorithm was used in both SOMs and MLPs, whilst in both LSVMs and DST-based classifier the leave-one-out cross (LOO) validation algorithm was adopted, also introduced in order to maximise the utilisation of the training cohort. Findings indicate that, for the classified patient dataset, the DST-based or “Cardiff Classifier” proved to be the most suitable method for classifying patients with knee OA due to its highest out-of-sample classification accuracy amongst the classifiers tested. This clinically relevant result partially validates the Cardiff classifier as a reliable diagnostic tool to assess the knee joint function of patients affected by knee OA. The DST-based classifier’s user-friendly interface lends itself to further development into a valuable clinical tool, enabling a clinician to objectively characterise knee function without requiring any assistance from biomechanists. There is also future potential to assist with patient assessment and surgical decision making, aimed at optimising patient outcome. Future work is underway to ascertain whether the Cardiff classifier is a clinically suitable technique for improving the accuracy of the diagnosis and monitoring the effectiveness of the prognosis related to other lower limb pathologies. 5. List relative to the paper: B, C, D, E, F and G. 6. My previous institution Cardiff University and colleagues working with Artificial Intelligence (AI)-based classifiers, as well as the International Society of Biomechanics (ISB) on July, 15th, 2015, thankfully judged my contribution to the field of Clinical Biomechanics as replicating at least equal results with respect to the ones published by Barton and Lees in different instances (1993, 1995, 1996, 2000), regarded as the International 'gold-standard' for neural network (NN) applications to gait analysis, but innovating in that I applied Lagrangian Support Vector Machines (LSVMs) to discriminate between healthy and osteoarthritic gait patterns and produced the first ever preliminary validation of the Cardiff Artificial Intelligence Classifier for aiding diagnosis of knee OA. If the Cardiff Classifier were used as an assistive AI-based tool for aiding diagnosis of knee OA, it may contribute to reduce the incidence of knee OA as well as pave the way to find a novel rehabilitative gait-based treatment of early knee OA, thus improving the life of numerous patients, especially considering our aging population worldwide more and more prone to being affected by knee OA. 7. Full citation of the paper: Parisi, L.; Biggs, P. R.; Whatling, G. M.; Holt, C. A. (2015). A novel comparison of artificial intelligence methods for diagnosing knee osteoarthritis. Presented in the special session no. 0027 and amongst the first five finalists of the Clinical Biomechanics Award at: 25th Congress of the International Society of Biomechanics, Glasgow, UK, 12-16 July 2015. 8. Any prize awarded in money would be equally divided amongst the four authors of this paper. 9. We expect this paper to be the 'best' by virtue of the outstanding contribution that we believe to have given to the Biomechanics and scientific community in general in extending the use of AI models to aid diagnosis of knee OA, thus contributing to reduce the incidence of knee OA as well as pave the way to find a novel rehabilitative gait-based treatment of early knee OA, improving the life of numerous patients, especially considering our aging population worldwide. 10. LCS (learning classifier systems), in particular supervised Neural Network (NN) model named "Multi-Layer Perceptron" (MLP) and unsupervised NN model named "Self-Organising Map" (SOM), the Machine Learning (ML)-based Lagrangian Support Vector Machine (LSVM) and the Dempster-Shafer-theory-based "Cardiff Classifier".