Prediction of Self-management Strategies for Future Quality of life and Health Statuses in Cancer Survivors Using Machine Learning Techniques: A Prospective Cohort Study with a Web-Based Survey

doi:10.21203/rs.3.rs-1983565/v1

Download PDF

Research Article

Prediction of Self-management Strategies for Future Quality of life and Health Statuses in Cancer Survivors Using Machine Learning Techniques: A Prospective Cohort Study with a Web-Based Survey

https://doi.org/10.21203/rs.3.rs-1983565/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Purpose: First, we aimed to develop and validate simple prediction models for QoL as a primary outcome and health statuses as secondary outcomes considering the use in actual clinical practice. Second, we aimed to interpret the prediction models using explainable artificial intelligence (XAI) methods by extracting important features and dependence plots. Third, we aimed to predict and interpret individual outcomes and visualize the results using the SHapley Additive explanation (SHAP) XAI method. This study's ultimate goal was to provide customized healthcare services for cancer survivors to improve their QoL and health statuses.

Methods: In this prospective cohort study with a web-based survey, we established the prediction model for QoL and health statuses and compared the performance with ensemble methods such as decision trees, random forest, gradient boosting, eXtreme Gradient Boost (XGBoost), and LightGBM. After the model comparison, we selected the XGBoost model for further analysis. We demonstrated critical features of the QoL and each health status separately and extracted the individual prediction result in the XGBoost model using SHAP.

Results: After data preprocessing and model selection, the final data set contained 256 cancer survivors with 38 predictive features. On repeated stratified K-fold validation, the XGBoost predictive model for the QoL showed high performance. The performance of the XGBoost predictive model for each health status (mental, social, and spiritual) was also good, except for physical health status. The important features of the above predictive models were different. In addition, our study showed a specific method of providing personalized healthcare services using the individual prediction results with SHAP.

Conclusions: This study was the first to develop and validate predictive models for the QoL and health status in cancer survivors, interpret the models, and provide a customized healthcare service with XAI technology.

Implications for Cancer Survivors: Our results can contribute to opening a new horizon in providing personalized healthcare services for cancer survivors with a web-based method.

SHAP

XGBoost

Quality of life

health status

cancer survivors

prediction model

Cancer survivors are continuously increasing in South Korea and around the world [1–3]. In South Korea, the recent 5-year cancer survival rate was 70.7% (2015–2019), and the number of prevalent cases reached above 2 million in 2019 [3]. The number of new cancer cases worldwide was approximately 19.3 million, while the number of cancer-related deaths was approximately 10 million in 2020 [4]. The 5-year cancer prevalence was over 50 million cases worldwide [4]. Thus, in 2040, the cancer burden is expected to be 28.4 million cases globally, and the number of cancer survivors is expected to increase further [5].

The increase in the number of cancer survivors has also increased the interest in improving their quality of life (QoL) and various health statuses such as physical, emotional, social, and spiritual health statuses [6–11]. As cancer survivors often suffer from physical problems such as fatigue, pain, sleep disturbances, and constipation; psychological disorders such as depression and anxiety; social difficulties such as loss of employment; and existential, spiritual difficulties related to the meaning of life as they face death after a life crisis [6–9], their higher interest in improving QoL and health statuses is a natural phenomenon accompanying the maintenance of daily life and survival. Cancer requires continuous management like a chronic disease due to the continuation of various difficulties accompanying cancer as well as the management of the disease itself [12]. Thus, self-management strategies are necessary for cancer survivors to take care of their own health for improving QoL and physical, mental, social, and spiritual health statuses.

The Smart management strategies for health Assessment Tool (SAT) has been recently developed to help cancer survivors practice self-management continuously and overcome various difficulties that come with cancer [13]. The SAT was developed by extracting cancer survivor-tailored crisis-overcoming strategies based on interviews with cancer survivors and cancer experts and literature reviews from multidisciplinary perspectives on behavioral changes and crisis-overcoming strategies in psychology, education, business administration, healthcare, and medicine [13]. The strategies outlined by the SAT are aimed at overcoming health crises and facilitating positive growth that can be applied in a time-specific manner according to the treatment period and health behavior stages of cancer survivors [13]. The SAT includes the following 15 sub-strategies in three sets, namely, SAT-C (Core strategy), SAT-P (Preparation strategy), and SAT-I (Implementation strategy [13]. A previous study reported that the more SAT strategies used by the survivor, the better was the QoL and overall health status (physical, mental, social, and spiritual health statuses). However, since that study clustered SAT strategies, it could not examine which strategies were important features individually predicting the QoL or each health status.

To our knowledge, no previous study on this topic has been performed using machine learning models that predict quality of life or health status, even though QoL and health status are important variables in predicting cancer survivorship [14–16]. Therefore, in this study, we intend to develop and validate a QoL and health status prediction model for cancer survivors that can be easily applied in actual clinical settings to improve the survivors’ QoL. To facilitate actual clinical application, this study attempted to simply organize the input variables of the machine learning prediction model into sociodemographic and clinical variables and SAT strategies for cancer survivors. Thus, this study identified the following three detailed research objectives with the ultimate goal of providing customized healthcare services for cancer survivors to improve their QoL and health status: (1) Development and validation of models to predict QoL primarily and overall health statuses (physical, mental, social, and spiritual) secondarily for cancer survivors; (2) explaining the prediction of the above model using explainable artificial intelligence (XAI) techniques and extracting important features that predict QoL and overall health statuses; and (3) predicting the QoL and overall health statuses of individual cancer survivors using XAI techniques and visualizing them to suggest a method for providing customized healthcare services.

Study Design

In this prospective cohort study with web-based surveys, participants completed the survey on the “HealthingU” website at baseline and after 6 months. We developed the HealthingU website to deliver survey results immediately with visually organized health information (Supplementary Information (SI) 1). Participants were encouraged to complete the surveys on the website, but those who found it difficult to take an online survey received the survey as well as the results of the survey by mail.

Study Participants

We included cancer survivors who met the following inclusion criteria: 1) diagnosed with cancer, 2) over 18 years of age, 3) able to read and write Korean, 4) able to use Internet and email, and 5) submitted written informed consent. A total of 540 cancer survivors participated in this study after submitting informed consent and completing a web-based survey with a coordinator’s support in outpatient clinics from four hospitals in South Korea. Among these 540 cancer survivors, 256 cancer survivors who completed the second survey after 6 months were ultimately included in this study. This study was conducted after obtaining approval from the institutional review boards of the four hospitals.

Measurements

Our questionnaire included questions regarding sociodemographic and clinical variables (age, sex, marital status, residence, income, education, religion, cancer stage, cancer type, and treatment stage), the SAT [13], the European Organization for Research and Treatment of Cancer (EORTC) core QoL questionnaire 30 (EORTC-QLQ-C30) [17], and the Health Status Questionnaire (HSQ), which evaluated physical, mental, social, and spiritual health statuses [18, 19].

The SAT, which measures self-management strategies for health in cancer survivors, includes three questionnaires: SAT-C, SAT-P, and SAT-I. SAT-C measures core self-management strategies for health with four sub-strategies: 1) positive-reframing, 2) proactive problem-solving, 3) experience-sharing, and 4) creating empowered relationships. SAT-P measures preparation strategies to support the preparation of healthy behavior practices or cancer treatment with five sub-strategies: 1) rational decision-making, 2) goal and action setting, 3) healthy environment development, 4) life value pursuing, and 5) priority-based planning. SAT-I measures implementation strategies to support the implementation of health behaviors and the treatment with six sub-strategies: 1) self-motivating, 2) activity-coping, 3) self-implementing, 4) energy-conserving, 5) reflecting, and 6) self-sustaining [13]. SAT scores ranged from 0 to 100. High SAT scores indicated that the cancer survivors used the strategies well. The difference in SAT scores between baseline and post-baseline surveys after 6 months was the secondary score minus the primary score.

The EORTC-QLQ-C30 is a representative questionnaire to measure cancer survivors’ QoL. The EORTC-QLQ-C30 measures various factors such as function, symptoms, and global QoL [17]. In this study, we only used the global QoL score as a factor representing cancer survivors’ QoL. Previous studies showed that global QoL was the most frequently reported factor for predicting cancer survivorship [15, 16]. The global QoL score also ranges from 0 to 100. High scores indicate that the cancer survivor has a high QoL. Global QoL was the primary outcome in this study.

The HSQ is a multidimensional questionnaire covering four domains of health status, namely, physical, mental, social, and spiritual health status [18, 19]. The HSQ measures health status with five-point Likert scales as follows: 1 = Excellent, 2 = Very Good, 3 = Good, 4 = Bad, and 5 = Very Bad. The health status was the secondary outcome in this study.

Data Preprocessing

Before developing and selecting the final model, data preprocessing was performed as follows:

First, we attempted to impute the missing data in the dataset. Since we used score data for SAT and EORTC-QLQ-C30, we performed missing imputation at the level of primary data acquired as Likert scales. Although various missing imputation methods are available, we applied the Markov Chain Monte Carlo (MCMC) multiple-imputation algorithm because it can handle an arbitrary pattern of missing data and can predict both categorical and continuous variables together [20]. For uniformity, we first eliminated the variables with more than 20% missing of all data. Thereafter, we applied MCMC to impute all the remaining variables.

Second, we converted all categorical variables into binary features for machine learning analyses, as shown in Tables 1 and 2. Only “cancer type” was converted into dummy variable using a one-hot encoding method because it was difficult to divide into binary features. We used the scores of continuous variables as it is in predictive models because we intended to propose a personalized healthcare plan for cancer survivors based on the scores of survey results showing how high and low their SAT scores are and the effect of the scores on their QoL and health status. Other preprocessing steps such as normalization and scalarization were not also performed since the ultimate purpose of this study was to provide customized health care services using individual cancer survivor information and SAT scores.

Third, outcome features in the dataset were encoded as binary values (Table 3). The primary outcome (global QoL) was encoded as high (> 33.3) or low (≤ 33.33) based on the scoring report, in which the global QoL score was classified in the problematic group when it was ≤ 33.33. The secondary outcome was encoded as “Good” (1 = Excellent, 2 = Very Good, 3 = Good) or “Bad” (4 = Bad, and 5 = Very Bad).

Fourth, we constructed a predictive model by removing a variable (the self-sustaining strategy in SAT) that showed high correlation (Pearson correlation > 0.7) with other variables to avoid multicollinearity problems and increase the stability of the model [21, 22].

Machine Learning Techniques

Various machine learning algorithms were used to create predictive models and compare their performance. In this study, we only used ensemble methods such as decision trees, random forest, gradient boosting, eXtreme Gradient Boost (XGBoost), and LightGBM. These ensemble methods did not require preprocessing such as scaling and normalization of each feature. We selected ensemble methods to use the scores of survivors’ survey results without any processing. We finally selected the XGBoost algorithm to develop predictive models and extract feature importance and individual prediction results.

Extreme gradient boosting

XGBoost is one of the most popular algorithms for tree-based ensemble learning [23]. In classification, XGBoost generally shows better prediction performance than other machine learning algorithms [24]. Although XGBoost is based on GBM, it is used more frequently than GBM because it addresses some of the limitations of GBM, including the slow execution time and lack of regularization [23, 25]. Although XGBoost and LightGBM are both relatively new GBM-based algorithms with no significant difference in predictive performance, LightGBM can easily show overfitting when applied to a small dataset [26]. Therefore, we finally selected XGBoost to develop the predictive models in this study.

Feature importance and individual prediction with Shapley additive explanations

Explainable artificial intelligence (XAI) techniques have recently emerged to increase the interpretability of complicated machine learning (ML) models [27]. Interpretability is critical in the medicine and healthcare fields to draw conclusions and make decisions based on the obtained results from MLT [28].

Measurement of Shapley Additive Explanation (SHAP) values is a well-known XAI technique. SHAP values can describe the “importance” of each feature to the overall model performance, with the effect’s direction indicated by positive or negative values [29]. SHAP values can also predict a particular data point (e.g., for one survivor) and explain why a particular data point showed a particular outcome (e.g., why one survivor had low QoL) [29]. The individual prediction is a new XAI function that could not be implemented using conventional statistical methods.

The SHAP value is obtained by calculating the net contribution for possible combinations of all variables and adding them [30]. To calculate the SHAP Value, the marginal contribution is measured while changing each variable by considering the combination of all variables. The average of marginal contributions obtained using this approach is called the SHAP Value [30]. SHAP values allow measurement of the contribution to the model by comprehensively considering the interactions between variables [30].

Model Development and Selection

In order to select a QoL predictive model (primary outcome), three XGBoost models were trained as follows:

Model 1: QoL predictive model based on sociodemographic and clinical variables

Model 2: QoL predictive model based on sociodemographic and clinical variables plus SAT scores

Model 3: QoL predictive model based on SAT scores alone

We compared these three models’ performance to determine the influence of SAT scores themselves on QoL predictions (SI 2). After identifying SAT scores’ prediction power, we ultimately selected Model 2, which included all individual personal characteristics and SAT scores together.

Final Model Training and Evaluation

To obtain the predictive models for QoL and health statuses, XGBoost models were trained using the Model 2 input features described above. The prediction performance of the XGBoost model was compared with four ensemble classifiers in MLT: decision tree, random forest, gradient boosting, and LightGBM, of which the first two are conventional classifiers and the other two are recent MLT classifiers.

All analyses were carried out using open-source libraries such as scikit-learn in Python 3.8. For the training dataset, we performed 5-fold cross-validation to verify optimal model hyperparameters by using a grid search. The optimal hyperparameters were applied on the test dataset to the test area under the receiver-operating-characteristic curve (AUROC) [28]. After hyperparameter tuning for each model, repeated stratified K-fold (K = 5, repeats = 20) validation was performed to evaluate the final models with the fine-tuned parameters (Fig. 1), wherein the models were trained on 4 folds iteratively and tested in the 5th fold [31]. Since the datasets were imbalanced, we chose to use repeated stratified K-fold validation for the evaluation.

Individual prediction, feature importance assessments, and visualized interpretations were carried out using open-source libraries such as scikit learn, shap, and xgboost in Python 3.8. We trained a gradient-boosted regression tree (GBRT) model using the open-source xgboost package (version 0.81). SHAP values were calculated using the open source shap package (version 0.29.1). We generated visualizations and graphs using this shap package. The area under the precision-recall curves (AUPRC), the AUROC, accuracy, and F1 scores were used for prediction model assessment [28].

Study Participants and Features

A total of 256 cancer survivors were included in the dataset after preprocessing. Tables 1 and 2 show the mean or prevalence of sociodemographic and clinical features and SAT score features. The mean age of the survivors was approximately 51 years; 64.5% of the survivors were women; 51.2% earned more than 4,000 won per month; 62.1% had an undergraduate degree or higher; 60.2% lived in a metropolitan area; 85.5% were married; 63.3% had a religion; 36.3% had breast cancer, 27% had lung cancer, 22.3% had colon cancer, 6.6% had gastric cancer, and 7.8% had other cancers; 53.9% had stage Ⅲ or Ⅳ cancers; and 66.4% completed treatment more than 5 years ago. The mean SAT baseline and change scores are described in Table 2.

Table 1 Sociodemographic and clinical features in the dataset
Features	Category	N (%) or Mean ± SD	Min-Max
Age		51.28 ± 9.64	23-77
Sex	Male	91 (35.5)
Sex	Female	165 (64.5)
Income (won)	<4,000,000	125 (48.8)
Income (won)	≥4,000,000	131 (51.2)
Education	≤High school graduates	97 (37.9)
Education	≥University graduates	159 (62.1)
Residence	Other areas	102 (39.8)
Residence	Metropolitan area	154 (60.2)
Marriage	Single	37 (14.5)
Marriage	Married	219 (85.5)
Religion	No	94 (36.7)
Religion	Yes	162 (63.3)
Cancer type	Breast cancer	93 (36.3)
	Lung cancer	69 (27)
	Colon cancer	57 (22.3)
	Gastric cancer	17 (6.6)
	etc.	20 (7.8)
Cancer stage	I, II	118 (46.1)
Cancer stage	III, IV	138 (53.9)
Treatment stage	≤5 years after treatment	86 (33.6)
Treatment stage	>5 years after treatment	170 (66.4)

SD, standard deviation

Table 2 Smart management strategies for health Assessment Tool (SAT) features in the dataset
Features	Mean ± SD	Min-Max
Baseline
SAT-C
Proactive problem-solving strategy	64.1 ± 18.93	10-100
Positive-reframing strategy	68.42 ± 21.44	14.8-100
Creating empowered relationship strategy	77.06 ± 18.73	11.1-100
Experience-sharing strategy	48.78 ± 25.19	0-100
SAT-P
Goal and action setting	48.68 ± 21.19	0-100
Rational decision-making strategy	63.54 ± 17.74	22.2-100
Healthy environment-creating strategy	57.32 ± 19.81	0-100
Priority-based planning strategy	54.69 ± 21.22	0-100
Life value-pursuing strategy	61.12 ± 20.68	6.7-100
SAT-I
Self-motivating strategy	57.63 ± 19.49	4.8-100
Self-implementing strategy	54.13 ± 21.58	0-100
Reflecting strategy	44.79 ± 27.16	0-100
Energy-conserving strategy	58.42 ± 19.27	0-100
Activity-coping strategy	57.58 ± 22.70	0-100
Change
SAT-C
Proactive problem-solving strategy	-1.18 ± 16.84	-100-36.7
Positive-reframing strategy	-1.35 ± 18.03	-91.6-51.9
Creating empowered relationship strategy	-4.32 ± 17.38	-94.4-38.9
Experience-sharing strategy	-1.74 ± 24.28	-88.9-66.7
SAT-P
Goal and action setting	0.13 ± 19.62	-90-53.3
Rational decision-making strategy	-3.08 ± 17.10	-88.9-55.6
Healthy environment-creating strategy	-0.62 ±18.09	-60-46.7
Priority-based planning strategy	-1.04 ± 19.83	-75-58.3
Life value-pursuing strategy	-3.33 ± 18.82	-73.3-46.7
SAT-I
Self-motivating strategy	-1.10 ± 17.99	-76.2-61.9
Self-implementing strategy	-0.42 ± 19.79	-75-58.3
Reflecting strategy	-3.32 ± 27.98	-100-100
Energy-conserving strategy	-0.95 ± 20.78	-77.8-20.78
Activity-coping strategy	-0.31 ± 21.38	-86.7-73.3

SD, standard deviation

Table 3 shows the outcome features in the dataset, with 88.67% showing a high global QoL (primary outcome), 66.41% showing good physical health status, 82.03% showing good mental health status, 88.67% showing good social health status, and 75.39% showing good spiritual health status (secondary outcome).

Table 3 Outcome features in the dataset
Features	Category	N (%)
Primary outcome
Global Quality of Life	High	227 (88.67)
Global Quality of Life	Low	29 (11.33)
Secondary outcomes
Physical health status	Good	170 (66.41)
Physical health status	Bad	86 (33.59)
Mental health status	Good	210 (82.03)
Mental health status	Bad	46 (17.97)
Social health status	Good	227 (88.67)
Social health status	Bad	29 (11.33)
Spiritual health status	Good	193 (75.39)
Spiritual health status	Bad	63 (24.61)

Model Performance

Table 4 shows the comparison results between XGBoost and other ensemble classifiers for the primary outcome (global QoL). XGBoost produced the best results for the main performance measures, i.e., the AUROC and AUPRC. Accuracy and F1 scores of XGBoost yielded the second-best results. Assessment of the prediction performance of XGBoost showed an AUROC of 0.80 (95% CI, 0.78 to 0.82), AUPRC of 0.96 (95% CI, 0.95 to 0.97), and F1 scores of 0.88 (95% CI, 0.84 to 0.92) and 0.93 (95% CI, 0.91 to 0.95). The final tuned hyperparameters are described in SI 3.

Table 4 Comparison of the prediction performance for primary outcomes in different methods
Methods	AUROC (mean, 95% CI)	AUPRC (mean, 95% CI)	Accuracy (mean, 95% CI)	F1 score (mean, 95% CI)
Decision Tree	0.63 (0.61-0.65)	0.91 (0.90-0.92)	0.88 (0.83-0.90)	0.94 (0.92-0.95)
Random Forest	0.78 (0.76-0.80)	0.96 (0.95-0.97)	0.88 (0.87-0.90)	0.94 (0.93-0.95)
Gradient Boosting	0.78 (0.76-0.80)	0.95 (0.94-0.96)	0.87 (0.81-0.92)	0.93 (0.89-0.95)
XGBoost	0.80 (0.78-0.82)	0.96 (0.95-0.97)	0.88 (0.84-0.92)	0.93 (0.91-0.95)
LightGBM	0.79 (0.77-0.81)	0.96 (0.95-0.97)	0.89 (0.88-0.90)	0.94 (0.94-0.95)

AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CI, confidential interval.

Table 5 shows the prediction performance results for health statuses (secondary outcome) using XGBoost. Overall, the predictive performance for the secondary outcome was slightly lower than that for the primary outcome. The predictive results for physical health status included an AUROC of 0.66 (95% CI, 0.65 to 0.67), AUPRC of 0.77 (95% CI, 0.75 to 0.79), accuracy of 0.69 (95% CI, 0.57 to 0.78), and F1 score of 0.79 (95% CI, 0.72 to 0.84). The predictive results for mental health status included an AUROC of 0.71 (95% CI, 0.69 to 0.73), AUPRC of 0.90 (95% CI, 0.89 to 0.92), accuracy of 0.83 (95% CI, 0.78-0.86), and F1 score of 0.90 (95% CI, 0.88 to 0.93). The predictive results for social health status included an AUROC of 0.77 (95% CI, 0.75-0.79), AUPRC of 0.96 (95% CI, 0.95-0.97), accuracy of 0.88 (95% CI, 0.84 to 0.92), and F1 score of 0.94 (95% CI, 0.92 to 0.96). The predictive results for spiritual health status included an AUROC of 0.75 (95% CI, 0.74 to 0.76), AUPRC of 0.87 (95% CI, 0.85 to 0.89), accuracy of 0.79 (95% CI, 0.71 to 0.88), and F1 score of 0.87 (95% CI, 0.81 to 0.92).

Table 5 The prediction performance for health statuses' outcomes in the XGBoost model
Outcomes	AUROC (mean, 95% CI)	AUPRC (mean, 95% CI)	Accuracy (mean, 95% CI)	F1 score (mean, 95% CI)
Physical Health Status	0.66 (0.65-0.67)	0.77 (0.75-0.79)	0.69 (0.57-0.78)	0.79 (0.72-0.84)
Mental Health Status	0.71 (0.69-0.73)	0.90 (0.89-0.92)	0.83 (0.78-0.86)	0.90 (0.88-0.93)
Social Health Status	0.77 (0.75-0.79)	0.96 (0.95-0.97)	0.88 (0.84-0.92)	0.94 (0.92-0.96)
Spiritual Health Status	0.75 (0.74-0.76)	0.87 (0.85-0.89)	0.79 (0.71-0.88)	0.87 (0.81-0.92)

AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CI, confidential interval.

The prediction performance of AUROC and AUPRC are depicted in figures 2 and 3 and SI 4. Figure 2 shows the AUROC and AUPRC for predicting the primary outcome (global QoL) in the XGBoost model, and Figure 3 shows the AUROC and AUPRC for predicting the secondary outcomes (overall health statuses) in the XGBoost models. SI 4 shows the AUROC and AUPRC of different algorithms for predicting the global QoL to compare the performance results with XGBoost.

Feature Importance and Individual Prediction

We applied the final tuned XGBoost models to extract SHAP values in each primary and secondary outcome predictive model. Figures 4-5 showed the feature importance bar plots and the beeswarm plots for the top ten important features in each model. For predicting the global QoL, the top three important features were the activity-coping strategy, change values of the self-implementing strategy for 6 months, and the proactive problem-solving strategy. For predicting the physical health status, the top three important features were the activity-coping strategy, the healthy environment-creating strategy, and age. For predicting the mental health status, the top three important features were the proactive problem-solving, the positive-reframing, and the rational decision-making strategies. For predicting social health status, the top three features were the activity-coping, the healthy environment-creating, and the self-motivating strategies. For predicting spiritual health status, the top three features were the positive-reframing strategy, religion, and income. Thus, the top three most important features for each predictive outcome were different. The baseline SAT scores were found to be more important than the change scores of the SAT strategy for 6 months in predicting outcomes.

We selected two samples using the global QoL model and examined the compositions of individual predictions. Blue arrows indicate that the variables contributed to decreasing the global QoL, whereas red arrows indicate that the variables contributed to increasing the global QoL. In Figure 6(A), the predicted global QoL was 2.07, suggesting that the global QoL was high (true) for this survivor, although the survivor was living alone and the low baseline score of the rational decision-making strategy in SAT-P had a negative influence on the QoL. In Figure 6(B), the predicted QoL was -2.58, suggesting that the QoL was low (false) for this survivor. Although there are some positive effects on the QoL, such as no reduction of using the self-implementing strategy for 6 months, the low baseline score of the proactive problem-solving strategy in SAT-C, the activity-coping strategy in SAT-I, the self-motivating strategy in SAT-I, and the reduction in using the activity-coping strategy for 6 months had more negative effects on the global QoL. As a result, this survivor’s global QoL turned out to be low (false).

In this study, we developed and validated various models for predicting cancer survivors’ QoL primarily and overall health statuses such as physical, mental, social, and spiritual health statuses secondarily. The QoL of cancer survivors was highly predictable only with basic personal characteristics such as sociodemographic and clinical variables and the scores of self-management strategies (SAT strategies) (AUROC = 0.80, AUPRC = 0.96). The development of a simple model for predicting QoL can increase the ease of application to improve the QoL of cancer survivors in clinical settings.

The most important feature predicting the QoL of cancer survivors was the activity copying strategy of the SAT-I. Previous studies have described the direct and indirect influence of the activity copying strategy. Popa-Velea et al. reported that the activity coding strategy has a significant influence on QoL through resilience [32], and other studies have shown that coping strategies are directly beneficial to the QoL and psychosocial adaptation of cancer survivors [33, 34]. In this study, among the 15 strategies of SAT, the activity-coping strategy was found to have the greatest influence on the QoL of cancer survivors. Thus, active usage of coping strategies can be recommended for cancer survivors to improve their QoL.

The second important variable was the change value of the self-implementing strategy in SAT-I. This is an indicator of the usage of the self-implementing strategy by cancer survivors for six months, and usage of this strategy more than the baseline was important in predicting the QoL. The self-implementing strategy is a strategy for implementing the health behaviors and treatment behaviors necessary for health determined by cancer survivors [13]. The use of the self-implementing strategy may affect health behaviors, which might be beneficial to QoL [13]. The third important variable was the proactive problem-solving strategy of SAT-C. As mentioned in the background of this study, cancer survivors experience various physical, mental, social, and spiritual difficulties more than the general population [6-10]. Therefore, the use of the proactive problem-solving strategy to actively address these problems might have a positive effect on the QoL.

Predictive performance of health statuses was modestly lower than that of the QoL in cancer survivors. Except for the physical health status predictive models, the predictive models for mental, social, and spiritual health statuses’ showed good performance: all AUROCs were over 0.70 and all AUPRCs were above 0.87. In a previous study, when SAT strategies were clustered and physical health status was compared between the groups with high and low strategy scores, the group with the high strategy score also showed a high average physical health status score. However, this study showed that the performance of SAT scores in classifying the goodness of physical health status was not very good. Perceived health status and QoL are distinct constructs [35]. Although physical, mental, social, and spiritual health status constitute a part of quality of life, the predictive model for SAT strategies and personal characteristics predicted QoL better than all health statuses. The QoL has a muti-dimensional nature combining various factors such as physical, psychological, social health statuses, and subjective evaluation of environmental facets surrounding the human beings [36]. Interestingly, our simple model with SAT scores predicted the QoL determined by these complex factors well. In addition, important features for the QoL and each health status appeared differently, suggesting that a customized self-management strategy considering each survivor’s QoL or health status is necessary when providing healthcare services for cancer survivors.

The ultimate goal of this study was to construct evidence for provision of customized healthcare services to cancer survivors and thereby improve their QoL. To apply the abovementioned feature importance results to cancer survivors, we could identify critical score points of important SAT strategies using SHAP values. On the basis of these critical score points, cancer survivors can be recommended to use strategies based on clear goals by presenting accurate numerical improvements (Figure 7). For example, Figure 7 depicted the influence of the top three important features on predicting the global QoL. Figure 7 (A) showed that the activity-coping strategy has a loose linear relationship with the QoL. A baseline activity-coping strategy score close to 45 or higher (the critical score point) has a positive effect on the QoL. Moreover, the QoL can be improved by educating and training SAT strategies for cancer survivors on the basis of the simple model we developed.

This study attempted to develop a web-based questionnaire targeting cancer survivors. Our team hoped that this approach would provide useful health information to cancer survivors while allowing research. To achieve that goal, the HealthingU website was developed, which allowed survivors to view personalized survey results about their health as soon as they completed the web-based survey. In this study, we found a way to provide truly customized healthcare information by using SHAP, a newly developed XAI technology [29, 30]. SHAP's individual prediction is a new function that is unavailable in existing traditional statistical methods, and this technology will open up a new horizon in providing customized healthcare services for cancer survivors. However, the interpretation of statistical significance of SHAP values was impossible in the context of ML models due to the absence of an established system [37]. Future studies should aim to accumulate data examining whether the results of individual predictions of SHAP have statistically significant positive effects on the QoL or health status of cancer survivors by using verifiable research methods such as randomized controlled trials (RCTs). If the effectiveness of individual prediction results of SHAP can be verified through several clinical studies, this approach will be more useful for cancer survivors in clinical practice.

This study had several limitations. First, the number of participants in this study was relatively small, and the findings should be evaluated in a larger number of samples to determine if the results of this study are reproducible. To alleviate the issues associated with the small sample size, we conducted hyperparameter tuning and repeated K-fold validation to reduce overfitting [38, 39]. In addition, we also evaluated the XGBoost model using bootstrap validation method (n_iteration = 300) [28]. The prediction performance after the bootstrap validation (AUROC = 0.77, AUPRC = 0.97) was similar to the predictive result of the repeated K-fold validation (SI 5). Second, the completion rate for the second survey completion rate was approximately 47% of the completion rate for the first survey. The first survey was conducted in outpatient clinics with a coordinator’s support. The second survey was answered online by the survivor alone, while cancer survivors who did not respond to the survey online within the period completed the survey by mail. At the second survey, most of the survivors completed the survey by mail. Since the average age of the participants in this study was 51.3 years old, it would have been difficult for older people to complete the web-based questionnaire alone. Support seems necessary to increase the web-based survey completion rate of the elderly. In addition, the user interface and user experience of the web-based survey may have to be customized to increase the ease of its use for the elderly. Third, the interpretation of SHAP values’ statistical significance is impossible in the current context of MLTs [37]. The efficacy of the individual prediction results should be verified by conducting a pilot test to examine that the individual prediction results in this study can have more statistically significant positive effects on cancer survivors’ QoL or health status.

The purpose of this study was to propose an approach for customized healthcare services for cancer survivors by developing and validating simple prediction models for QoL primarily and health status secondarily by producing important features predicting the health outcomes, and by implementing individual predictions. Using MLTs such as XGBoost, we obtained good prediction results for QoL and health statuses. The XAI techniques, e.g., SHAP in this study, enabled us to verify important features for each outcome with a high stable prediction potential and to identify a new way to apply customized healthcare services for cancer survivors. These study results can be helpful for promoting cancer survivors’ QoL and health status and developing customized healthcare plans in clinical settings.

Conflicts of Interest: None declared

Funding: The Ministry of Health & Welfare in South Korea (1320330)

Author Contributions: All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Ju Youn Jung. The first draft of the manuscript was written by Ju Youn Jung and Young Ho Yun commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Data Availability: The data that support the findings of this study are available from the corresponding author upon reasonable request.

Ethics approval: This study was conducted after approved by the institutional review boards of the four hospitals in South Korea.

Consent to participate: Informed consent was obtained from all individual participants included in the study.

Miller, K.D., et al., Cancer treatment and survivorship statistics, 2016. CA: a cancer journal for clinicians, 2016. 66(4): p. 271-289.
Miller, K.D., et al., Cancer treatment and survivorship statistics, 2019. CA: a cancer journal for clinicians, 2019. 69(5): p. 363-385.
Kang, M.J., et al., Cancer statistics in Korea: incidence, mortality, survival, and prevalence in 2019. Cancer Research and Treatment: Official Journal of Korean Cancer Association, 2022. 54(2): p. 330.
Ferlay, J., et al., Global cancer observatory: cancer today. International Agency for Research on Cancer. Lyon, France, 2020.
Sung, H., et al., Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians, 2021. 71(3): p. 209-249.
Kang, D., et al., Impact of objective financial burden and subjective financial distress on spiritual well-being and quality of life among working-age cancer survivors. Supportive Care in Cancer, 2022. 30(6): p. 4917-4926.
Lee, E.S., et al., Health-related quality of life in survivors with breast cancer 1 year after diagnosis compared with the general population: a prospective cohort study. Annals of surgery, 2011. 253(1): p. 101-108.
Shin, H.W., et al., Correlates of existential well-being and their association with health-related quality of life in breast cancer survivors compared with the general population. Breast cancer research and treatment, 2009. 118(1): p. 139-150.
Chochinov, H.M., Depression in cancer patients. The lancet oncology, 2001. 2(8): p. 499-505.
Stark, D.P.H. and A. House, Anxiety in cancer patients. British journal of cancer, 2000. 83(10): p. 1261-1267.
Lee, M., et al., Employment status and work-related difficulties in stomach cancer survivors compared with the general population. British Journal of Cancer, 2008. 98(4): p. 708-715.
McCorkle, R., et al., Self‐management: Enabling and empowering patients living with cancer as a chronic illness. CA: a cancer journal for clinicians, 2011. 61(1): p. 50-62.
Yun, Y.H., et al., Patient‐reported assessment of self‐management strategies of health in cancer patients: development and validation of the Smart Management Strategy for Health Assessment Tool (SAT). Psycho‐Oncology, 2015. 24(12): p. 1723-1730.
Greer, J.A., et al., Perceptions of health status and survival in patients with metastatic lung cancer. Journal of pain and symptom management, 2014. 48(4): p. 548-557.
Montazeri, A., Quality of life data as prognostic indicators of survival in cancer patients: an overview of the literature from 1982 to 2008. Health and quality of life outcomes, 2009. 7(1): p. 1-21.
Montazeri, A., et al., Quality of life in lung cancer patients: as an important prognostic factor. Lung cancer, 2001. 31(2-3): p. 233-240.
Yun, Y.H., et al., Validation of the Korean version of the EORTC QLQ-C30. Quality of life research, 2004. 13(4): p. 863-868.
Yun, Y.H., et al., Employee health behaviors, self-reported health status, and association with absenteeism. Journal of occupational and environmental medicine, 2016. 58(9): p. 932-939.
Yun, Y.H., et al., Attitudes of the general population, cancer patients, their family caregivers, and physicians toward dying and death: a nationwide survey. Global Journal of Health Science, 2017. 9(10): p. 201-201.
Kim, S. and D. Kim, Imputation method for missing data based on clustering and measure of property. The Korean Journal of Applied Statistics, 2018. 31(1): p. 29-40.
Zeynalova, A., et al., Preoperative evaluation of tumour consistency in pituitary macroadenomas: a machine learning-based histogram analysis on conventional T2-weighted MRI. Neuroradiology, 2019. 61(7): p. 767-774.
Kocak, B., et al., Machine learning-based unenhanced CT texture analysis for predicting BAP1 mutation status of clear cell renal cell carcinomas. Acta Radiologica, 2020. 61(6): p. 856-864.
Chen, T. and C. Guestrin. Xgboost: A scalable tree boosting system. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016.
Putatunda, S. and K. Rama. A comparative analysis of hyperopt as against other approaches for hyper-parameter optimization of XGBoost. in Proceedings of the 2018 International Conference on Signal Processing and Machine Learning. 2018.
Li, B., et al., Genomic prediction of breeding values using a subset of SNPs identified by three machine learning methods. Frontiers in genetics, 2018. 9: p. 237.
Kinnander, M., Predicting profitability of new customers using gradient boosting tree models: Evaluating the predictive capabilities of the XGBoost, LightGBM and CatBoost algorithms. 2020.
Gunning, D., Explainable artificial intelligence (XAI), Defense Advanced Research Projects Agency (DARPA), 2017. 2017.
Domínguez-Olmedo, J.L., et al., Machine learning applied to clinical laboratory data in Spain for COVID-19 outcome prediction: model development and validation. Journal of medical Internet research, 2021. 23(4): p. e26211.
Lundberg, S.M. and S.-I. Lee, A unified approach to interpreting model predictions. Advances in neural information processing systems, 2017. 30.
Lundberg, S.M., G.G. Erion, and S.-I. Lee, Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888, 2018.
Haider, S.P., et al., Prediction of post-radiotherapy locoregional progression in HPV-associated oropharyngeal squamous cell carcinoma using machine-learning analysis of baseline PET/CT radiomics. Translational oncology, 2021. 14(1): p. 100906.
Popa-Velea, O., et al., Resilience and active coping style: Effects on the self-reported quality of life in cancer patients. The international journal of psychiatry in medicine, 2017. 52(2): p. 124-136.
Nipp, R.D., et al., The relationship between coping strategies, quality of life, and mood in patients with incurable cancer. Cancer, 2016. 122(13): p. 2110-2116.
Tuncay, T., Coping and quality of life in Turkish women living with ovarian cancer. 2014.
Smith, K.W., N.E. Avis, and S.F. Assmann, Distinguishing between quality of life and health status in quality of life research: a meta-analysis. Quality of life research, 1999. 8(5): p. 447-459.
Organization, W.H., WHOQOL-BREF: introduction, administration, scoring and generic version of the assessment: field trial version, December 1996. 1996, World Health Organization.
Li, R., et al., Machine learning–based interpretation and visualization of nonlinear interactions in prostate cancer survival. JCO Clinical Cancer Informatics, 2020. 4: p. 637-646.
Hu, C.-A., et al., Using a machine learning approach to predict mortality in critically ill influenza patients: a cross-sectional retrospective multicentre study in Taiwan. BMJ open, 2020. 10(2): p. e033898.
Brownlee, J., XGBoost With python: Gradient boosted trees with XGBoost and scikit-learn. 2016: Machine Learning Mastery.

No competing interests reported.

220813SupplementaryMaterial.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Prediction of Self-management Strategies for Future Quality of life and Health Statuses in Cancer Survivors Using Machine Learning Techniques: A Prospective Cohort Study with a Web-Based Survey

Status:

Version 1

Abstract

Figures

Introduction

Methods

Study Design

Study Participants

Measurements

Data Preprocessing

Machine Learning Techniques

Extreme gradient boosting

Feature importance and individual prediction with Shapley additive explanations

Model Development and Selection

Final Model Training and Evaluation

Results

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1