LOGISTIC REGRESSION, RANDOM FOREST AND NEURAL NETWORKS TO PREDICT TKR POSTOPERATIVE PAIN

J.-M. Blasco Igual¹, S. Roig-Casasús¹, E. Soria-Olivas², A. Serrano¹, C. Sanchez¹, B. Díaz¹, J. Pérez-Malétzki¹, Y. Alakhdar¹, D. Hernández-Guillén¹

¹Universitat de València, Physiotherapy, Valencia, Spain, ²Universitat de València, Ingeniería Electrónica, Valencia, Spain

Background: Total knee arthroplasty is a prevalent procedure to decrease pain in advances stage of the condition, when conservative treatments fail to resolve clinical symptoms. Traditional regression models have been used to predict clinical outcome, based on preoperative evaluations of a number of bio-psycho-social parameters. However, these models have limitations in the adjustments, so using more advanced data analysis methodologies based on neural networks and machine learning processes can help to understand the importance of the mechanisms that lead to better clinical results. Among them, the influence of exercise and the number of supervised physiotherapy sessions to reduce pain are aspects that need further study and are of special interest to us. However, most of data analysis procedures are still understudied whereas evidence that validated the models is still scarce.

Purpose: The goal was to test different machine learning algorithms to study the importance of rehabilitation exercises on the reduction of reported pain 6 months after the procedure of total knee arthroplasty.

Methods: A cohort study that included 253 patients undergoing primary total knee replacement. Machine learning prediction models logistic regression, random forest and neural networks were train to predict the decrease or not decrease of the pain after total knee arthroplasty. The decrease or not of the pain was set as the target variable. The pain before the intervention and variables of physical exercise and rehabilitation exercises were used as predictive variables. A 80:20 split was applied to the study population in order to create a training set and an independent test set. For each model, fine tuning was done and the area under the ROC curve (AUC) and confusion matrix were calculated using the test set.

Results: The best model was a neural network with one hidden layer made of two neurons, activation relu, solver sgd, learning rate constant and alpha equal to 0.0001. For our test set, it gave us an AUC of 0.8235 and a confusion matrix indicated that the model correctly predicted 17 out of 17 cases of decrease of pain and 5 out of 7 cases of not decrease of pain on patients not used in the training set. We computed the importance of the variables and the results suggested that the most important aspects in the clinical outcome were the pain reported before total knee arthroplasty, to exercise after the intervention and the number of rehabilitation days completed by the patient.

Conclusions: A successful prediction was obtained for the decrease or not of pain and the variables which were the most predictive were the pain before the intervention, if the patient exercised and for how long they did rehabilitation.

Implications: This study supports that the level of preoperative pain and life style are good indicators of postoperative outcome.

Funding acknowledgements: This project PID2020-115825RA-I00, was funded by MCIN/ AEI/10.13039/501100011033, convocatoria Proyectos I+D+i 2020 - Modalidades “Retos Investigación” y “Generación de Conocimiento”

Keywords:
Total knee replacement
Machine learning
Pain

Topics:
Musculoskeletal: lower limb
Innovative technology: information management, big data and artificial intelligence
Orthopaedics

Did this work require ethics approval? Yes

Institution: Hospital La Fe de Valencia

Committee: Comité de Ética del Hospital La Fe de Valencia

Ethics number: 2018/280