Machine Learning Algorithm Guiding Local Treatment Decisions For Lung Cancer Patients With Bone Metastases


 Background: As life expectancy increases for lung cancer patients who develop bone metastases, the need for personalized local treatment for bone metastases is expanding.Methods: Lung cancer patients with bone metastases were treated by a multidisciplinary team via surgery, percutaneous osteoplasty, or radiation. The pre- and post-treatment visual analog scale (VAS) and Quality of Life (QoL) scores were analyzed. QoL at 12 weeks was the main outcome. Treatment-related costs and overall survival time (OS) were collected. We used machine learning to develop and test models to predict which patients should receive local treatment. Models discrimination were evaluated by the area under curve (AUC), and the best one was used for validation in clinical use. Results: Under the direction of a multidisciplinary team, 161 patients in the training set, and 32 patients in the test set underwent local treatment. A decision tree model included VAS scale, bone metastases character, Frankel classification, Mirels score, age, driver gene, aldehyde dehydrogenase 2, and enolase 1 expression had a best AUC of 0.92 (95%CI 0.89 to 0.94), and 36 patients in a validation set underwent local treatment guided by the model. Improved QoL and VAS scores were observed at 12 weeks after local treatment in training, test, and validation sets (p < 0.05), with no significant differences among the three datasets. There were no significant differences in mean costs among the three datasets in the four treatment groups. OS was 18.03±0.45 months and did not significantly differ among treatment groups or the three datasets. Conclusions: Local treatment not only had no negative influence on OS but also provided significant pain relief and improved QoL. QoL, OS or costs did not significantly differ between patients whose treatment was guided by a multidisciplinary team or machine learning model. Our machine learning model using clinical data can help guide clinicians to make local treatment decisions to improve patients’ QoL.Trial registration: No. ChiCRT-ROC-16009501

Spinal stability was ascertained using Spinal Instability Neoplastic Scores (SINS) [13], and the risk of pathological fracture for the appendicular skeleton was ascertained using the Mirels scoring system [14]. Surgical procedures followed guidelines of the Global Spine Tumor Study Group and Italian Orthopedic Society [15][16][17][18]. Procedures for POP were introduced by our MDT in 2012 [7]. Radiation was performed mainly with 6-MV photons using linear accelerators. Dose fractionation schedules included multi-fraction radiation, such as 30 Gy in 10 fractions. Adjuvant therapy-like radiation [19] was used after surgery or POP to prevent tumor recurrence.
Informed consent was obtained for all patients in the study. If local treatment was performed, informed consent by the patient or a legal guardian was obtained 24 hours before initiation and after thorough explanation of the methods, potential complications, and alternative treatments.
Patients were asked to complete a questionnaire that assessed severity of pain using the visual analog scale (VAS) [21] and QoL using the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Bone Metastases Module (EORTC QLQ-BM22) [22][23] 1 day before and at 1, 6, 12, and 24 weeks after local treatment or enrollment. QoL at 12 weeks was the main outcome. Patients were followed for survival every 3 months.

Cost Valuation
Cost analyses of individual patients were estimated from a payer perspective using health resource utilization data from patient charts. Costs of procedures performed during an inpatient stay were assumed to be captured in diagnosis-related group costs.

Model Development
Decision trees (DT) (eXtreme Gradient Boosting, XGBoots), support vector machine (SVM), and Bayesian neural networks (BNN) were used to build local treatment decision-making models. Based on our previous research [12], we selected the following predictor variables: sex, age, ECOG score, VAS score, bone metastases character, extent of bone metastases, visceral metastases, Frankel classi cation, and Mirels scale. As additional predictor variables, we selected lung cancer pathology, lung cancer driver gene, and ve differentially expressed bone metastasis proteins. The model output was: 0, no local treatment; or 1, local treatment. The training set including patients who had improved VAS and QoL measures after local treatment.

Model Performance
In the training set, we used 10-fold cross validation. The test set consisted of data not associated with the training set. Model discrimination was evaluated by area under the receiver operator characteristic curves (AUC). Sensitivity, speci city, and accuracy were used to evaluate model performance.

Model Validation
In the validation set, we used the model to make a decision regarding local treatment. The MDT made the nal decision about which local treatment to provide to patients and could reject the model's decision.

Statistical Analysis
Stata Corp 2013 (Stata Statistical Software: Release 13; StataCorp LP, College Station, TX, USA) and Python Version 3.6 (Python Software Foundation, Wilmington, DE, USA) were used to analyze data and build the model. Median values and ranges were determined for descriptive statistics. Chi-square and Fisher's exact tests were used for categorical variables. Student's t-tests and Mann-Whitney tests were used for continuous and ordinal variables. Wilcoxon signed-rank tests were used to compare paired outcomes at various follow-up times. The Kaplan-Meier method was used to estimate survival. A p < 0.05 was considered statistically signi cant.

Demographic and Clinical Characteristics
We enrolled 746 patients: the training set included 513 patients enrolled from October 24, 2016 to June 30, 2018; the test set included 108 patients enrolled from July 1, 2018 to October 31, 2018; and the validation set included 125 patients enrolled from November 1, 2018 to February 25, 2019. Of these, 161 patients in the training set, 32 patients in the test set, and 36 patients in the validation set underwent local treatment. A ow chart of the study is shown in Fig. 2A. Patient demographics and clinical characteristics did not signi cantly differ among the three datasets as shown in Table 1. Treatments in training, test, and validation sets are shown in Fig. 2B-D. .69 Post-Treatment Pain VAS scores before treatment for all 746 patients in surgery, POP, radiation, and no local treatment groups were 5.70 ± 1.22, 5.53 ± 1.34, 6.62 ± 1.48, and 3.37 ± 1.38, respectively; scores were highest in the radiation group and lowest in the no local treatment group (p < 0.05). VAS scores in surgery, POP, and radiation groups decreased signi cantly to 4.78 ± 1.28, 4.37 ± 1.36, and 5.39 ± 1.31, respectively, at 12 weeks after local treatment (p < 0.05). VAS scores for patients in the no local treatment group did not signi cantly differ 12 weeks after enrollment. Detailed scores are showed in Fig. 3A.
VAS scores in training, test, and validation sets all decreased signi cantly at 12 weeks after surgery, POP, or radiation (p < 0.05), with no signi cant differences among the three sets. Detailed scores are showed in Fig. 3B-D.

Post-Treatment QoL
Pain sites (PS) and pain characteristic (PC) scores of the QLQ-BM22 before treatment for all 746 patients were highest in the radiation group and lowest in the no local treatment group (p < 0.05), while functional interference (FI) and psychosocial aspects (PA) scores were highest in the no local treatment group and lowest in the radiation group (p < 0.05). Patients had improved QoL scores 12 weeks after surgery, POP, or radiation (p < 0.05). PS and PC scores decreased signi cantly while FI and PA scores increased signi cantly at 12 weeks after local treatment in surgery, POP, and radiation groups (p < 0.05). PS, PC, FI, and PA scores for patients in the no local treatment group did not signi cantly differ 12 weeks after enrollment. Pre-treatment and posttreatment subscores in pain and functional domains in QLQ-BM22 are shown in Fig. 3A.
In training, test, and validation sets, PS and PC scores decreased signi cantly while FI and PA scores increased signi cantly at 12 weeks after surgery, POP, or radiation (p < 0.05), with no signi cant differences among the three sets. Detailed scores are showed in Fig. 3B

Discussion
Management of bone metastases has been considered palliative and unassociated with patient prognosis and thus has not been given much importance in the past. Recently, however, it has become necessary to initiate bone management programs concurrently with cancer treatment to effectively prevent serious complications of bone metastases. Mechanical stability, neurological risk, oncological parameters, and preferred treatment (MNOP) algorithms [25] have been used to manage bone metastases since 2017. The algorithm used at our center since 2016 accounts for tumor histology (EGFR mutation is considered an indicator for good prognosis), tumor burden (extent of bone metastases and visceral metastases), patient performance (life expectancy), and technical di culty (complication).
Similar to previous studies, surgery, POP, or radiation provided signi cant pain relief and improved QoL [6,[26][27][28][29][30]. The MNOP algorithm suggests surgery or radiation as the main treatment for spinal metastases. Minimally invasive approaches such as POP are also recommended by our algorithm. We have treated hundreds of spinal PF and instabilities with POP instead of high-risk surgery. We nd that patients recover spinal stability after POP with low morbidity, and our results here show that mean costs for spine metastases in the POP group were much lower than the surgery group. However, we note that POP does not easily restore structural integrity and weight-bearing for the appendicular skeleton and that surgery can quickly restore these functions with less risk. To prevent PF, we have used preventive surgery or POP, and our algorithm shows that surgery and POP are complementary. We prefer POP for spinal metastases and surgery for appendicular skeleton metastases. In this study, we found that our algorithm was effective in providing treatment decisions that provided signi cant pain relief and improved QoL.
Our results show that local treatment did not negatively in uence OS for lung cancer patients with bone metastases. According to some studies, patients surgically treated for bone metastases survive < 10 months [27]. Recently, Tang et al reported OS of 14 months in lung cancer patients with spinal metastases, and patients who underwent surgery had longer survival [28]. However, our study excluded patients with a life expectancy of < 3 months, 92.6% patients received systemic medical treatments (57.9% for targeted agents), and Frankel classi cation in the no local treatment group was almost E, while in Tang's study it was A-D, possibly accounting for the longer survival in our study.
Machine learning models can help guide treatment decisions. Although the MDT approach is an effective method to manage bone metastases, it can be di cult to manage patients who may develop serious SREs in a timely manner by holding weekly meetings. Alternatively our machine learning models based on routinely available clinical parameters were constructed for local treatment decision-making. XGBoost is a novel boosting tree-based ensemble algorithm that has gained wide popularity in the machine learning community [31]. The DT model showed greater accuracy than SVM and BNN models, and it included driver genes and that ALDH2 and ENO1 expression had higher accuracy, which is in line with current needs for precision medicine. For ethical reasons, the nal decision was still left to the MDT. However, QoL, mean cost, and OS of the four treatment groups did not signi cantly differ between MDT and DT model decisions. Feasibility, stability, and economic e ciency of the DT model were satisfying, and the DT model was good at determining whether patients should receive local treatment. Thus, this tool may help clinicians decide on local treatment for individual patients.
To the best of our knowledge, this is the rst attempt to use machine learning techniques for local treatment decision-making models in lung cancer patients with bone metastases. However, the algorithm cannot completely solve the problem of patient classi cation, and the machine learning model could only guide decisions about whether to apply local treatment or not. However, as more patients receive local treatment, thus increasing data availability, we will be able to develop additional models that can better guide types of local treatment. However, some treatments have not been carried out in our study, which represents a weakness. For example, stereotactic body radiation therapy for painful spine metastasis shows better results in local control and pain relief than standard 2D or 3D techniques [30], and recent development of immune checkpoint inhibitors has fundamentally changed how patients with metastatic lung cancer are treated [5]. Further, this study is limited in that it involved a single center. Multi-center trials would provide additional evidence, although standardizing and homogenizing use of local treatment in different centers are challenging problems that remain to be solved.

Conclusions
Local treatment not only had no negative in uence on OS but provided signi cant pain relief and improved QoL in patients in our study. There were no signi cant differences between MDT and DT model decisions in QoL, mean costs, and OS for local treatment patients. Our machine learning model using clinical data can help guide clinicians to make local treatment decisions to improve patients' QoL.

Consent for publication
All authors have approved to publish this manuscript.

Availability of data and materials
The datasets generated and/or analysed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Competing interests
There are no nancial or other relationships that might lead to a con ict of interest. The authors declare that they have no competing interests. Authors' contributions WZY and SJ: study concept, design, analysis, and manuscript drafting. SY, GYF, XYM, ZBZ, and LYH: local treatment perform. YMD, YGY, and ZYY: patient followup, data collection. DDP and ZH: study concept, design, and manuscript editing. All authors have read and approved the nal manuscript.