Visual analogue scale foot and ankle vs. short-form 36 quality of life scores: artificial intelligence using machine learning analysis with an external validation

C. Angthong, P. Rajbhandari, W. Angthong

Faculty of Medicine, King Mongkut’s Institute of Technology Ladkrabang, Bangkok, Thailand. chayanin.an@kmitl.ac.th

---

• Artificial intelligence
• Orthopedics

OBJECTIVE: We aimed to utilize artificial intelligence (AI) via machine learning (ML) to analyze the relationship between visual analogue scale foot and ankle (VASFA) and short-form 36 (SF-36) quality of life scores and determine AI’s performance over the aforementioned analysis.

MATERIALS AND METHODS: We collected data from our registry of 819 data units or rows of datasets of foot and ankle patients with VASFA, SF-36 scores, and other demographic data. They were prepared and verified to be a proper input for building ML models using a web-based algorithm platform. After the first ML model was developed using random forest regression, the SF-36 percentage value was set as an endpoint. We developed a second ML model to evaluate it against the current algorithm. This new model employed a gradient-boosting regressor, where we omitted a key parameter, SF_Total, to correct the overfitting. We performed an external validation based on an unseen dataset from 42 data units of patients.

RESULTS: Internal validity showed an excellent relationship among the VASFA, SF-36 total score, and overall SF-36 percent values at a correlation coefficient (R2 score) of 1.000 based on the random forest regression model of ML (first model: 28XJ). The VASFA percent value of the total score (0=worst; 100=best) demonstrated the dynamic changes in the three zones of the score levels; these were unsatisfactory: ≤ 57.25; borderline: 57.26-80.99; satisfactory: ≥ 81 and could impact the levels of overall SF-36 percent value. A second ML model (model FK13) showed an R2 score of 0.977, which was a great performance. External validation showed no significant difference between the predicted and actual values, with a two-tailed p-value of 0.2136.

CONCLUSIONS: Our ML models predicted excellent relationships among VASFA, with or without SF-36 total score and overall SF-36 percentage values, with evidence from external validation.

Free PDF Download