Background

Cervical cancer is one of the most common cancers worldwide [1]. The disparity between the mortality rates amongst the high and the low income group countries signifies a health inequity [1]. Further, there exists a higher prevalence in the low socio-economic groups within different countries, which exemplifies uneven availability and accessibility to the health services [2].

Situation

The age-adjusted incidence of cervical cancer in India is 25.9% (1, 34,420 incident cases), which is higher than the average for South East Asian region [3]. According to the Population Based Cancer Registries (PBCR), it has the second highest prevalence following breast cancer [4]. The pervasiveness of cervical cancer is much higher in rural and low socioeconomic groups in India [2], [5]. Lack of access to screening, affordability, compliance and follow up of treatment are the major causes for this pattern of discordance [6].

Economic Burden in India

In addition to the significant contribution to the mortality rates, cervical cancer also leads to the loss of productive life due to prolonged disability [1]. It contributes to considerable economic burden, as the women of age group 25 - 64 years tend to be the sole caretakers of their families and households; and in some cases, they as well contribute to the family income [7]. Additionally, they are further deprived due to high medical costs, especially since most of the cases in developing countries are diagnosed at later stages,when the treatment is costly combined with poor prognosis [8]. Accordingly, there is a need to prevent the soaring secondary costs and economic burden as the more cost effective treatment modality can effect the outcome of patient’s post treatment quality of life.

The current study compares the scores of pre-treatment and post-treatment quality of life. The quality of life has been documented to vary with the treatment, the time since diagnosis and the cancer site [9]. Additionally, there are cultural differences which affect the quality of life assessment perception, as there are dissimilar beliefs regarding the constituents of normality and disease [10]. Being diagnosed with the disease is associated with social stigma, due to the belief that it is caused by sexual promiscuity, poor hygiene and the use of oral contraceptives [11]. Given that these beliefs are different from the Western world, therefore, reasonably there should be differences in the perception of quality of life in India. Thus, we conducted the study in patient population of north India and determined whether there is any significant difference in the post treatment quality of life. Furthermore, we structured a prediction model based on the findings of our study, which would serve to forecast the quality of life.

Ethics Statement

Institutional approval was obtained after reviewing the study protocol by the Institutional Review Board of the Himalayan Institute of Medical Sciences and Research, Uttarkhand. Written informed consent were obtained from patients after clear and concise explanation about the study. All the women in this study were informed about the right to refrain from participation in the study without any affect on the quality of health care services being provided to them. Patient confidentiality was assured by coding the patients’ information and removing the identifiable personal data before the data analysis.

Study Population

All patients who had undergone surgical treatment for cervical cancer between April, 2007 to March, 2011 at the tertiary care academic hospital, Himalayan Institute of Medical Sciences and Research, Uttarkhand in Northern India were selected for this study. Of the 361 patients who gave completely filled the consent form, 108 were excluded due to non-availability of post treatment assessment data. 55 subjects were excluded due to accompanying psychiatric disease, cognitive impairment, diagnosis of multiple cancers, ambiguity in determining follow-up time or incomplete information, and Stage IV of cancer. After exclusion, 198 subjects with cervical cancer were included in the study group.

Measures and Instruments

The patients were interviewed using two structured questionnaires-Health Related Quality of Life (HRQoL) questionnaires (EORTC QLQ C-30 and CX-24), to collect information about socio-economic and clinical status. The baseline values were recorded when the patient first reported to the hospital (T1) and second evaluation was done at a follow-up visit 6 months post treatment (T2). Clinical evaluation was performed at the same time points and relevant information was documented.

The European Organisation for Research and Treatment of Cancer (EORTC)- general cancer quality of life score questionnaire (QLQ C-30, and its cervical cancer module (QLQ CX-24) were used to measure HRQoL. These questionnaires have been extensively tested in multicultural and multidisciplinary settings, and have been confirmed to be reliable and valid [12]–[14]. The EORTC QLQ C-30 questionnaire comprises of 30 questions which assess functioning (physical, role, cognitive, emotional, social) and symptoms (fatigue, nausea and vomiting, pain, dyspnea, insomnia, appetite loss, constipation, diarrhea, financial difficulty), and a global health status score which assesses the overall QOL. The EORTC QLQ CX-24 questionnaire consists of 24 questions assessing functioning (body image, sexual enjoyment, sexual/vaginal functioning) and symptoms (symptoms experience, lymphoedema, peripheral neuropathy, menopausal symptoms, sexual worry). Both questionnaires use a four-point response scale (not at all, a little, quite a bit, and very much) to assess each functional or symptom item, and a seven-point response scale is used to assess global health status (from very poor to excellent). For model development the categorical raw scores were linearly transformed into a score of 0 - 100 for processing according to the EORTC manual [15].

Systemic literature review, established clinical knowledge and by performing univariate linear regression analysis following features were found to be statistically significant and were used in the development of the prediction model: age, marital status, vaginal bleeding, vaginal discharge, dyspareunia, abdominal pain, weight loss, parity, difficulty in controlling bowels or emptying bladder, cervical cancer stage, treatment, lymphoedema and peripheral neuropathy.

Prediction Model Development

For the development of the prediction model we compared the performance of 4 different algorithms. Support Vector Machines(SVM) was utilized with two different kernels- Linear and Radial Basis Function(RBF), Artificial Neural Network(ANN) and regularized Logistic Regression(LR). The preprocessing of the dataset was performed to standardize and normalize the features before applying learning algorithms. The quality of life questionnaire mean results were obtained in continuous data format on a scale of 0 - 100, which was converted into discrete 3 class output for each scale, to train the classifier machine learning algorithms (See Appendix S1). All the 13 variables were provided to the models and trained separately for functional, symptom and global health/QoL scales. The Scikits-Learn module in Python programming language was utilized for the development of Logistic Regression and SVM models [16]. The ANN multilayer perceptron(MLP) was developed using PyBrain module in Python [17].

Prediction Model Validation and Comparison

Cross-validation method with stratified K-Fold iterator was employed to derive a reliable estimate of the performance. Cross-validation approach splits the whole data several consecutive times in different train and test set, and returns the averaged value of the prediction scores. Stratified K-Fold return group of samples, called “folds” by preserving the same percentage for each target class as in the complete set. The stratified 6-folds were implemented to estimate the accuracy. Predicted values were then combined across the 6 runs and summarised by mean Area Under Curve(AUC). The mean accuracy, mean area under the curve (AUC),mean squared error (MSE) and adjusted-for-chance mutual information index (AMI) of each of the models were calculated for performance comparison between the Logistic Regression, SVM (‘Linear’), SVM (‘RBF’) and ANN models. An individual cervical cancer patient was the unit of analysis in this study. The area under the ROC (Receiver Operating Characteristics) curve or simply AUC has been validated to evaluate the ranking performance of the machine learning algorithms [18]. The MSE, which is computed as difference between the true and predicted values and then averaged across data, was used as an indicator of effectiveness with model fits [19]. Adjusted Mutual Information (AMI) is an adjustment of the Mutual Information (MI) score to account for chance [20](See Appendix S1).

Statistical Analysis

We used R, an open source statistical programming environment for univariate linear regression analysis; and scientific computing Python packages to generate logistic regression, SVM and ANN models. [16], [17], [21], [22]. The student’s test or test was utilized for comparing the characteristics in the study population and the quality of life scales before and after treatment .

Predictive Factors Significance

The significance of the predictive factors previously selected using univariate linear regression analysis were compared. Table 4 shows the coefficients of significant variables for Symptom scale, Functional scale and Global health/Quality of Life (GH/QoL) post treatment in the logistic regression model. All the selected variables were found to be statistically significant ().

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Coefficients(coef) of significant variables for Symptom scale, Functional scale and Global health/Quality of Life (GH/QoL) post treatment in Logistic Regression model, a Generalized linear model (GLM) type.

https://doi.org/10.1371/journal.pone.0089851.t004

Comparison of Prediction Models

Table 1 shows the performance comparison of four machine learning algorithms on Symptom, Global health/QoL and Functional scales for the prediction of post-treatment cervical cancer quality of life outcomes. This has been demonstrated in terms of Mean Squared Error (MSE), mean Area Under Curve (AUC), Adjusted for chance Mutual Information index (AMI) and accuracy for prediction.

Symptom scale.

Comparison of the four models reveal, that all the models performed suitably on the Symptom scale. On the basis of accuracy, 3 models Support Vector Machine with Linear kernel-SVM (Linear), Support Vector Machine with Radial Basis Function-SVM (RBF) and LR (Logistic Regression) scored ; and SVM (Linear) outperformed other models with accuracy of . In terms of mean AUC, SVM (Linear) and Artificial Neural Network (ANN) surpassed SVM (RBF) and LR (mean AUC = 0.90, 0.85, 0.80 and 0.72, respectively). The mean squared error (MSE) was low for all four models varying between 0.020.03. SVM (Linear) was superior to other models SVM (RBF), LR and ANN in terms of AMI (AMI = 0.90, 0.81, 0.82 and 0.82, respectively). (Table 5, Figure 1A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Mean Area Under Receiver Operating Characteristic (ROC) curve.

Mean ROC (AUC) for Support Vector Machine algorithm with Linear Kernel for (A) Prediction of Symptom scale was 0.90. (B) Prediction of Global Health/QoL was 0.84. (C) Prediction of Functional scale was 0.85.

https://doi.org/10.1371/journal.pone.0089851.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The performance comparison of four machine learning algorithms on symptom, global health/QoL and functional scales for the prediction of post treatment cervical cancer quality of life outcomes.

https://doi.org/10.1371/journal.pone.0089851.t005

Development of Prediction Model

Based on these findings, we developed a prediction model tool for estimating the post treatment HRQoL outcome on Functional, Symptom and Global Health/QoL scales. During the development of the model, we compared the well established four machine learning algorithms, among which SVM(Linear) proved to be most accurate and consistent. To the best of our knowledge, this study is the first one to develop a prediction model for the post-treatment QoL. It also depicts that, given the same clinical and socio-demographic inputs; and output target classes, the predictive accuracy of SVM(Linear) is the highest and most consistent for HRQoL scales.

Recently, SVM and ANN models have been used for non-linear modelling in diverse fields, ranging from bioinformatics [38]–[40], neurosciences [41], [42], imaging [43], [44] to clinical prediction model development [45], [46].

We implemented univariate linear regression analysis during feature selection stage and validated our features by logistic regression coefficients values. Logistic regression was prioritized over linear regression due to multi-class nature of the classification. Cross validation with stratified K-Fold method was implemented to rule out varying percent of target class in a test set. This practice prevented biased predictions, over-fitting and hence provided with the accurate cross validation scores. We utilized AMI rather than Normalized Mutual Information index (NMI), as NMI fails to account for chance in comparing clustering performances by supervised learning algorithms [20] (See Appendix S1). The prediction model aims to determine the quality of life in an Indian context, which is essentially different from the developed countries in terms of socio-sexual environment and financial context which have been important determinants, as found in the study.

Online Tool

We developed an online tool PrediQt-Cx, utilizing our prediction model, which is available at http://prediqt.org. The web application was developed using Python modules and back-end implementation of SVM (Linear) algorithm. The aim of developing the web application was to facilitate the accessibility of prediction model for critical evaluation, utilization for supporting treatment decisions and for external validation.

Limitations

Our study was limited to north Indian population only. Thus, the applicability of model will be limited until external validation in cross-cultural and diverse socioeconomic settings is undertaken. Despite the multiple options being provided as treatment modality, the training of models and predictions were based only on three alternative sets of models (surgical treatment modalities, radio-chemotherapy and multiple interventions). This was decided as the sample sizes for some treatment modalities (conization, total abdominal hysterectomy and chemotherapy) were small and model training on them was not possible without prediction bias. Similarly FIGO (International Federation of Gynaecology and Obstetrics) stages were clubbed together into four groups. Grouping was based on stage specific clinical treatment decisions to account for small sample sizes in some stages (Stage IA1 and Stage IA2).

Conclusion

In this study, the prediction model PrediQt-Cx, which was based on SVM, was developed and internally cross validated. After external validation, PrediQt-Cx can be employed in decision making procedure by clinicians and patients from north India region. It has been made available for open access at http://prediqt.org.