Description
BACKGROWND AND OBJECTIVES
Partial meniscectomy is the most popular surgical technique for the treatment of meniscal tears and its results have been widely studied. The surgical outcomes worsen with presence of osteoarthritis changes, meniscal root tears, complex tears and narrowing of the articular space. Failure of the procedure has been described over 30% at 5 years so the preoperative identification of patients with high risk of failure is mandatory1,2,3,4.
Medical application of big data analysis is gaining popularity, being useful for the design of predictive models since their algorithms can process big amount of data and identify trends in disease development567.
The main objective is to establish the foundations of an AI tool that assists orthopaedic surgeons in identifying patients at high risk of failure in meniscal surgery. As secondary objectives we mean to determine if the risk factors associated with surgery failure are preventable at a pre-, intra- or postoperative level.
METHODS
Retrospective multicentric study in wich all patients who underwent partial meniscectomy from January 2018 to February 2022 were included. Inclusion criteria were age over 45 years and first episode of meniscal surgery while exclusion criteria were traumatic tears and less than 12 months of follow up.
As demographic variables age, sex, BMI, type of job and nationality. Personal history of cardiovascular or oncologic disease, hypertension, diabetes, hyperlipidemia, chronic kidney disease, OSAS, Parkinson disease, fibromyalgia, depression or anxiety disorder were collected. Preoperative physical examination, radiographic and MRI findings were analyzed. As outcome variables we compiled a satisfaction survey, the Lysholm Knne Scoring Scale and the need of a second meniscal surgery
A total of 350 patients were identified, only 244 (56% male with a mean age of 54 years) answered the satisfaction survey and were included in the final data set analysis
RESULTS
Outcome analysis revealed a 73.8% of patients with residual knee pain with a 49% of VAS score over 5. 83.2% of the patients returned to work but 37.6% gave up physical activity because of knee pain.
Data augmentation technique was applied to increase the variability of the dataset resulting in a better representation of the target population avoiding the overfitting and improving the accuracy of the predictive tool. Synthetic data were obtained with two models, GAN and LSTM, resulting in a total of 5000 synthetic data. Decision tree, random forest, gradient boosted tree (GBT) and multilayer perceptron (MLP) were used as predictive algorithms of Lysholm score results.
The most accurate models turned to be the decision tree (82% accuracy and 7.95% error) and random forest (80% accuracy and 7.6% error). The best sensitivity for the detection of patients with high risk of failure (poor Lysholm score) corresponded to the decision tree (96% vs 91%) (FIGURE 5), while the greatest specificity was achieved in the random forest model algorithm (96% vs 89%). The combination of these parameters results in a F1 score of 92% form the decision tree and 94% for the random forest, being the random forest the best model for predicting the results of meniscal surgery using artificial intelligence.
CONCLUSION
Data-driven approach helps us predict the future by using past and current information. Having tools capable of adjusting the individual probability of failure of a surgery allows improving the quality of care by assisting in decision making.
