A gender-bias-mitigated, data-driven precision medicine system to assist in the selection of biological treatments of grade 3 and 4 knee osteoarthritis: development and preliminary validation of precisionKNEE

IntroductionOsteoarthritis is a leading cause of global disability and is set to worsen with the concurrent rise in rates of obesity and an ageing population [1]. Current clinical solutions are sub-optimal with regards to their invasiveness and outcomes. Orthopaedic biologics is an emerging field that offers alternative and parallel treatment options to address this problem. Determining which patients will benefit most from these novel treatments is key in developing clinical pathways. MethodsOur dataset included 329 patients treated with microfragmented fat injection (MFAT) over a 2 year period. Clinico-demographic data was recorded as well as 1-year Oxford Knee Score (OKS). The data was modelled to predict OKS 1-year response using Random Forest Regressors. Gender-bias was mitigated and outliers were hidden from the training model. The model was validated on raw test data and on a subset of patients with Kellgren-Lawrence grade 3 and 4 radiological evidence of arthritis, age greater than 64, preoperative OKS less than or equal to 27 and idiopathic aetiology of arthritis. ResultsThe mean age and mean body mass index (BMI) of patients in our dataset was 66.4 years, 26.9 respectively. 53.5% of patients had Kellgren-Lawrence grade 4 arthritis. The final models RMSE was 6.72, MAE was 5.38 and r-squared was 0.23 on raw test data. An RMSE of 9.77 and MAE of 7.81 was achieved when validating the model on our subset of patients. Wilcoxon signed rank tests found no evidence of predicted results being statistically significantly different to ground truth values (p {inverted question} 0.05). Preoperative OKS and Kellgren-Lawrence arthritis grade was the most important feature in our model. DiscussionOur model is performant and able to predict 1 year OKS response outcome within our set of patients. We have found key features of prediction and would recommend these are researched further to improve model performance. Our dataset does not compare outcomes with other standard treatments. We also dont compare outcomes with other biologic treatments. Ultimately, this research can be used as a tool to benefit both patients and clinicians in a combined decision-making process.