With rising life expectancy in cancer patients with bone metastases, the need for local treatment (LT) is expanding. Machine learning (ML) could create reasonable generalizations, the purpose of this article was to evaluate the use of ML model in LT strategies. Patients were treated by an interdisciplinary team in Shanghai Sixth People's Hospital. Visual analog scale (VAS) and Quality of Life (QoL) Questionnaire Bone Metastases Module scores were analyzed before, 1 week, 1, 3, and 6 months after treatments. ML models were used to calculate LT probability, and confusion matrix was used to calculate the accuracy, precision, recall and F1 score of models. ML models were further used to calculate pathological fracture (PF) probability in lung cancer patients. Of 386 patients enrolled between 2016 and 2017, 101 patients underwent LT. Significant improved VAS and pain domains scores were observed in 27 surgery patients at 1, 3 and 6m, while functional domains scores at 3 and 6m. All five scores improved significantly in 46 percutaneous osteoplasty patients at 1w, 1 and 3m. Significant improved VAS and pain domains scores were observed in 28 radiation patients at 1 and 3m, while functional domains scores at 3m. Compared with team decision-making, decision tree was superior to support vector machine and Bayesian neural networks in model building. The VAS scale, primary cancer, Frankel classification, Mirels score, C-terminal telopeptide of type I collagen (CTx), age, mid-fragment of osteocalcin (MID), character of bone metastases, CA153, and visceral metastases were included in the DT model. In 386 patients, the values of decision tree for the accuracy, precision, recall and F1 score were 86.53%, 78.44%, and 64.90% and 0.69 respectively. 124 lung cancer patients were used to calculate PF probability by decision tree. We also put driving gene mutation and five differentially expressed proteins into the model. The VAS scale, character of bone metastases, age, driving gene mutation, C-terminal telopeptide of type I collagen (CTx), mid-fragment of osteocalcin (MID) and enolase 1 (ENO1) were included in the DT model. The sensitivity, specificity and accuracy of DT model was 90.52%, 87.19% and 77.72%. Appropriate LT provided significant pain relief and improvement in QoL. The ML model is effective in helping physicians determine which patient may be the candidate for LT.

Citation Format: Hui Zhao, zhiyu wang, jing Sun, yifeng Gu, mengdi Yang, guangyu Yao. Exploiting machine learning in local treatment strategies in cancer patients with bone metastases: A real world clinical trial [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2059.