Exploring the role of health-related quality of life measures in predictive modelling for oncology: a systematic review
Auteurs:
T. G. W. van der Heijden, K. M. de Ligt, N. J. Hubel, S. van der Mierden, B. Holzner, L. V. van de Poll-Franse, B. H. de Rooij, the EORTC Quality of Life Group
Health related quality of life (HRQoL) is increasingly assessed in oncology research and routine care, which has led to the inclusion of HRQoL in prediction models. This review aims to describe the current state of oncological prediction models incorporating HRQoL. A systematic literature search for the inclusion of HRQoL in prediction models in oncology was conducted. Selection criteria were a longitudinal study design and inclusion of HRQoL data in prediction models as predictor, outcome, or both. Risk of bias was assessed using the PROBAST tool and quality of reporting was scored with an adapted TRIPOD reporting guideline. From 4747 abstracts, 98 records were included in this review. High risk of bias was found in 71% of the publications. HRQoL was mainly incorporated as predictor (78% (55% predictor only, 23% both predictor and outcome)), with physical functioning and symptom domains selected most frequently as predictor. Few models (23%) predicted HRQoL domains by other or baseline HRQoL domains. HRQoL was used as outcome in 21% of the publications, with a focus on predicting symptoms. There were no difference between AI-based (16%) and classical methods (84%) in model type selection or model performance when using HRQoL data. This review highlights the role of HRQoL as a tool in predicting disease outcomes. Prediction of and with HRQoL is still in its infancy as most of the models are not fully developed. Current models focus mostly on the physical aspects of HRQoL to predict clinical outcomes, and few utilize AI-based methods.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Plain English summary
The review lays the groundwork for the new field of using quality of life in prediction for oncology. It summarizes how quality of life is used in prediction models, identifying the factors of quality of life that are used. It also assesses the quality of reporting and model development of models including quality of life for cancer patients.
It has been found that the factors that associated with symptoms and physical functioning of a patient are used the most. Quality of life factors are mainly used as predictors, while a few models try to use quality of life to predict quality of life. Improvements are needed for reporting the development of models with quality of life, as the majority has a high risk of bias and are missing steps according to guidelines.
To summarize, the field of quality of life in prediction models has grown over the years, but is far from having a model used by patients and doctors in the clinic.
Introduction
Health related quality of life (HRQoL) can be measured using patient-reported outcome measures (PROMs), which are standardized questionnaires completed by patients that quantify the impact of the disease on their daily life and HRQoL [1, 2], and are used to monitor patients in routine clinical care and as a measure of quality of care in registries [3]. The integration of PROMs in clinical routine has rapidly increased in recent years [4] and is accompanied by a substantial increase of available data on patients’ HRQoL [5]. In its aggregated form, this data is currently mainly used in decision aids and as quality indicators of care [6].
HRQoL data could further improve individual care by its use in prediction modelling, which uses a patient’s available data to predict a prognostic or diagnostic outcome or event at a later time point [7]. In oncology, models that predict survival and treatment outcomes are at the forefront, like PREDICT and GAIL score [8‐10]. These prediction models and their outcomes are currently based on clinical and disease characteristics of patients and are used by clinicians to inform patients about their life expectancy or course of treatment [7, 9, 11, 12].
Integration of HRQoL in prediction models for oncology is a recent advancement, as the volume of collected HRQoL data has grown [13]. Although a comprehensive overview of model applications does not yet exist, examples include prediction of overall survival (OS) or progression free survival (PFS) by physical functioning and symptom scales [14, 15], and HRQoL outcomes can be predicted by treatment type or cancer characteristics [16, 17]. Fatigue at baseline is predictive for later global HRQoL, demonstrating HRQoL as predictor and outcome in the same model [18]. Furthermore, the choice of prediction model methods depends on study design, outcome, and type of data available [19]. Currently, there are two methodological approaches distinguishable in prediction modelling: Artificial intelligence (AI) and the more traditional “classical statistical” methods. As there can be overlap between methods (e.g. regression), the distinction lies in the aim of the prediction, while for classical methods the aim is inference, the aim of AI methods is to uncover patterns and hidden layers in the data regardless of causation. At this point, it remains unclear if AI has added value and if there is a preference for a prediction method for incorporating HRQoL data.
Whilst the field of HRQoL prediction modelling in oncology has expanded from proof-of-concept models to the first validated models [13], there is no published overview that summarizes this emerging field. The aim of this review is to provide a synthesis regarding which HRQoL domains play a role in the field of oncological predictions. The research questions that are addressed in this review are:
Which analysis and methods are used in oncological prediction that include HRQoL domains?
Which HRQoL domains are used as predictors, outcomes, or both in the field of oncological prediction?
What are the predictors’ relations with the outcome and vice versa?
Methods
This review was registered in PROSPERO (CRD42023456777). No protocol was developed for this review. The ‘transparent reporting of a multivariable prediction model for individual prognosis or diagnosis: checklist for systematic review and meta-analysis’ (TRIPOD-SRMA) was used for reporting this review [20]. PubMed, Embase via Embase.com, SCOPUS and CINAHL via EBSCO databases were systematically searched using search terms covering HRQoL, PRO(s) and PRO measures, development and validation of prediction models, and any type of cancer. Both MeSH terms and Emtree terms were used where appropriate. As no MeSH terms exist for prediction model related keywords, the search filter for prediction methods was based on Geersing et al. [21]. For the full search terms, see Appendix A. The search was performed in February 2023, and updated in January 2024.
For each publication, the abstract was screened using RAYYAN by two independent screeners (TvdH and either KdL or BdR) [22]. Conflicting decisions were judged unblinded by a third screener, whose judgment was final (KdL or BdR). For the abstract screening, the inclusion criteria were limited to studies that involved adult cancer patients aged 18 years or older; the study also had to report a statistical or AI-based model that includes HRQoL (functioning or symptoms) as an independent or dependent variable. Finally, the study had to be longitudinal with a minimum of two time points spaced at least three months apart, with at least one time point including HRQoL measurement. Publications about PROMs development and validation, case reports, conference abstracts, not in English, and literature reviews were excluded.
The full-text screening was done by two independent reviewers (TvdH, KdL, BdR) using RAYYAN. Conflicting decisions were judged unblinded by a third screener (KdL or BdR), who’s judgment was final. For the full-text screening, inclusion criteria were: the publication aimed to develop or validate a prediction model, full multivariate models were reported, and any performance measure (e.g. R2, C-statistic or ROC) of the model was reported.
Quality of reporting assessments of the publications were done by one independent reviewer (either TvdH or BdR) using an adapted version of the ‘transparent reporting of a multivariable prediction model for individual prognosis or diagnosis’ (TRIPOD) checklist [23] (See Appendix B). Publications that did not fulfil at least the majority (≥ 7 out of 13) of the, for this review, selected criteria were deemed of insufficient quality of reporting and were therefore excluded.
All remaining publications were eligible for data extraction. The risk of bias of the prediction models described in the remaining publications was assessed by one independent reviewer (either TvdH or BdR) using the ‘prediction model risk of bias assessment tool’ (PROBAST) [24], assigning certainty assessment to each publication to indicate the quality of the described models.
For each included publication, the study design, type of analysis, type of cancer, type of PROMs, HRQoL used as outcome or predictor, number of time points, type of performance measures and development/validation of model were extracted to report the overview of prediction models in HRQoL of cancer. No synthesis of the performance measures was performed as the focus of this review was on how HRQoL is used in prediction models. Data extraction was performed using a excel sheet by one independent reviewer (BdR or TvdH). Missing items from extraction were left as is.
Results
Record screening
The search yielded 10726 records; after de-duplication, 4747 records were eligible for the abstract screening. Of those, 645 were eligible for full text screening. Main reasons for exclusion at this stage were that studies did not apply a longitudinal study design (n = 1162), did not present a prediction model (n = 901), or included a population not affected by cancer (n = 366). After full text screening, 128 records remained for quality screening. Excluded records after full text screening did not report a full prediction model (n = 94), or did not report HRQoL in model (n = 215; Fig. 1).
Fig. 1
Flowchart of inclusion and exclusion during screening process
×
Quality of reporting
After quality screening, 98 records remained that were eligible for data extraction and risk of bias assessment (Fig. 1). From the included papers, 43% of the publications were not transparent or unclear in their reporting of how missing data was handled for development of a prediction model (Appendix B). In addition, a description how model performance was assessed was lacking in 24% of publications. For the results section, 86% of publications reported a full model, 75% reported model performance measures (e.g. R2, C-statistic or ROC). However, updating or validation of a model after initial development was only described in 10% of the records.
Risk of bias
The risk of bias of each study was assessed using PROBAST (Robert F. [24]). Only 23% of the studies were assessed as having a low risk of bias (RoB) (Appendix C). Reporting of predictors, outcomes, and population of the models was of sufficient quality. However, the poor reporting of the analysis was the main cause of the high RoB rating of 71% the studies, which included the use of univariate associations to select the variables for multivariate models, the lack of a missing data strategy, or model overfitting and optimism not being accounted for.
Emerging field of research
Of all included publications, the first was published in 2006; from 2013 on, the publications steadily increased over 10 years to 19 publications per year in 2023 (Appendix D). Of note is the introduction of the AI based models from 2020, which contributed to the accelerated increase of publications and thus interest in this field (Appendix D). The majority of the included publications (76%) was published in journals in the first quartile of the SCImago journal ranking (SCImago). 18% of publications were in Q2 ranked journals, 4% of publications in Q3, none in Q4. However, 2% of publications were in unranked journals.
Population
The most included patient populations were patients with lung (15%), breast (15%), head and neck (11%), or prostate (10%) cancer (Appendix E), whilst 7% of studies included all cancer types. The size of the populations for the models varied from 34 to 30.969 patients. The distribution is strongly right skewed with a median of 294 (177.5–891.5 IQR) patients (Appendix E).
Questionnaires
Nearly half of the studies used the European Organisation for Research and Treatment of Cancer (EORTC) QLQ-C30 (43%) to measure HRQoL, 9% used Functional Assessment of Cancer Therapy (FACT) and 3% Patient-Reported Outcomes Measurement Information System (PROMIS) questionnaires (Appendix E). For assessing and predicting emotional domains, the Hospital Anxiety and Depression Scale (HADS) (10%), Center for Epidemiologic Studies (CES)-Depression (2%) and Distress thermometer (2%) were frequently used. Of the disease-specific questionnaires, EORTC QLQ-LC13 (7%) and Expanded Prostate Cancer Index Composite-26 (EPIC-26) (6%) questionnaires were the most frequently used (Appendix E).
Time points of HRQoL measurement
Most studies measured HRQoL at two time points (51%); the median was 2 (IQR 2-4) with a maximum of 15 (Appendix E). In eight studies (8%), HRQoL was measured at one of two included time points, meaning that HRQoL was predicted by another type of input data or vice versa. Three publications did not report how often they measured HRQoL in the methods, yet from the models was derived that they had at least a minimum of two HRQoL time points.
Methods of prediction models
The majority of papers in this review described the use of classical statistical models (84%), mostly including Cox proportional hazard (CPH) regression (36%), linear (19%) and logistic regression (18%) while in 16% of papers AI models were used (Appendix E). For AI methods (16%), tree based methods like random forest (6%) and gradient boosting (6%) were mostly used, followed by neural networks (4%) (Appendix E). Of note is the trend that the AI papers were published from 2020 on, while the first classical model including HRQoL in this review is from 2006 (Appendix D).
Health-related quality of Life as predictor or outcome
In the majority of studies included in this review, HRQoL variables were included in the models as predictors (76%) (Table 1; Appendix E). Global QoL (26%) and physical functioning (21%) were most frequently included as predictors. Furthermore, emotional symptoms including anxiety/depression (13%), emotional functioning (8%) and role functioning (7%) were used as predictors. Other predictors included symptom scales such as fatigue (16%), pain (15%), dyspnoea (9%) and loss of appetite (8%).
Table 1
The summary overview of studies included in this review, displaying the main information per publication
Twenty-three studies (23%) included HRQoL as both predictor and outcome in the prediction model (Appendix E). Nearly all of them used a baseline or later score to predict HRQoL further in time. Three studies selected change in HRQoL as predictor for their models [25‐27].
HRQoL was used as a modelled outcome in 21 (21%) studies, with the domains of sexual functioning (9%), fatigue (7%), physical functioning (7%) and urinary function (7%) being the most commonly selected outcomes (Appendix E). Few models focused on summary score (6%) and global QoL score (5%), which is an aggregate score of (almost) all scales of the questionnaire. However, these outcomes are predicted by non-HRQoL variables like demographic characteristics, treatment type, dosage and location of treatment [28‐34], and (bio) markers [35].
Predictive relations
Almost 40% of the models that included HRQoL as predictor showed a predictive relation between HRQoL and survival, predicted through Cox regression analyses (OS, PFS) (Appendix E; Appendix G ). For instance, change in HRQoL score at disease progression predicted survival in lung cancer patients [36]; EQ-5D scores predicted survival in glioma patients [37], and physical functioning predicted survival in pancreatic cancer patients [38].
Other models focused on the relation between HRQoL and treatment. In the prostate cancer models, symptoms like urinary incontinence and irritation were predicted by the dose and location of radiotherapy [28, 31, 32, 34]. During chemotherapy, cancer patients could be discriminated and assigned to different HRQoL trajectories over time based on baseline HRQoL scores [29, 82]. Misra et al. [43] found that receiving chemotherapy was predictive of fatigue in breast cancer.
Subsequent HRQoL scores predicted by previous scores were frequently based on patients’ baseline scores. For example, functional problems were predicted by self-reported anxiety and/or depression scores [28], and baseline fatigue was predictive of fatigue at a later time point [61]. Only one study used all baseline HRQoL domains to predict the summary score of a PROM [64] (see Table S3).
Discussion
The aim of this review was to identify and synthesize the HRQoL domains that are most relevant to oncological predictions. In our review of 98 studies on HRQoL prediction in oncology, we found that three-quarters of the selected publications used HRQoL domains as predictors. In contrast, HRQoL was less commonly included as an outcome (21%), where it was primarily predicted by treatment. Physical functioning and symptoms were more frequently included in prediction models than role, cognitive and social functioning. Emotional functioning was often included as predictor, but mostly in models aiming to predict future HRQoL scores based on other HRQoL domains.
Methods of HRQoL prediction in oncology
AI versus classical models
In this review, no notable differences in variable selection were found between AI and classical statistical methods; similar outcomes and predictors were included in both. However, AI models represented only one eighth of the included publications and had been published more recently (2020-present). This may be explained by the recent increase in volume of available HRQoL data needed for AI-predictions and the general shift in focus towards AI and exploring AI opportunities [123]. Several publications compared multiple AI based strategies to explore which methods work to predict with HRQoL [34, 46]. For large datasets, AI based prediction methods are considered a better option than classical methods, as these methods better account for the complexity of the underlying data [124, 125]. Interestingly, in this review, sample sizes varied across all approaches and included only four large datasets (> 5.000 patients); several smaller sized studies (< 100, n = 13) appeared to be underpowered [93, 126]. In the current review, we noted that the majority of publications used AI methods in an explorative manner. Therefore, our findings do not facilitate a direct comparison of AI and classical methods, and thus, we cannot conclude whether AI methods outperform classical methods.
HRQoL dichotomization
We noted that HRQoL as outcome or predictor was either expressed in its original continuous format, dichotomized using a threshold score, or expressed as a change score (over time) (data not shown). While clinically relevant thresholds for impairment have been developed, just a few of the models took this approach [127]. This approach has upsides and downsides. Using a dichotomised variable over a continuous variable can make a model less accurate and requires a larger sample size [128]. Therefore, from a modelling perspective, continuous and change score modelling are preferred for HRQoL prediction modelling. However, binary scores might be preferred from a clinical perspective as these facilitate clinical interpretation: it clearly discriminates between patients who are in need of extra care and those who do not. We would argue that a cut-off can be applied on a continuous outcome score, thus balancing both clinical and modelling preferences.
Methodological limitations with current HRQoL prediction models
All current models have contributed to finding HRQoL predictors and outcomes that can be used in predictive modelling. Yet, other methods of modelling not mentioned in this review could also be used. Importantly, dynamic models using joint modelling approaches were not reported in this review [129]. This is a major gap in both knowledge and implementation, as clinical settings continually generate new data over time through repeated HRQoL assessments and routine clinical tests. Dynamic models, which update predictions as new information becomes available, more accurately reflects the real-world situation in the clinic [130, 131].
Furthermore, the quality of reporting and bias assessments reveal significant gaps in current model development, highlighting simple yet crucial steps needed for improvement in this field.
A crucial first step is to accurately assess model performance which was not done or reported in a substantial proportion of the models (90%). Assessing model performance both in a training- and test dataset (i.e. internal validation) is essential to ensure model validity and to apply corrections like penalization and shrinkage, reducing overfitting and model optimism.
Secondly, many models either did not report a missing data strategy or only addressed it for part of the data types (e.g. only for HRQoL). Some publications applied complete case analysis, which is valid only if sensitivity analysis confirm that included and excluded groups are similar. We believe that using the full dataset, even if imputed, is preferable to complete case analysis as it minimizes overfitting and optimism by utilizing all available data.
Development standards for prediction models
The majority of the solutions to the issues identified in the models in this review are outlined in guidelines for development, validation and reporting of prediction models [23].Few papers included in this review explicitly mentioned the use of the TRIPOD standard for developing or reporting prediction models; even though this checklist was published in 2015 and 90 (92%) included publications were published after 2015. The quality check using TRIPOD criteria revealed that the level of the reporting was deemed poor. Notably, most of the AI-based models had higher reporting quality compared to the publications using classical statistical models based on the TRIPOD criteria. This difference in scoring might be due to the nature of the AI models, as they follow standard development procedures that include the majority of the selected TRIPOD criteria like model performance and validation. In contrast, classical models are also used to find possible predictors based on association rather than correlation. Therefore, these do not require rigorous development standards, because the aim of the research is not the complete development of a model.
Since the publication of the TRIPOD checklist in 2015, several additional TRIPOD guidelines have been published, like TRIPOD-AI for AI models and TRIPOD-Cluster for clustered data [132, 133]. Thus, to improve model development, transparency and reporting quality we strongly urge new studies in the field of HRQoL prediction to use the available TRIPOD checklists.
HRQoL domains in oncological prediction models
HRQoL as predictor
Fatigue and physical functioning domains and baseline scores for all sorts of domains were most often applied as predictors, regardless of model type, outcome or cancer type. This selection may have been driven by the recommendations about important outcomes for cancer research by the FDA [134]. A large number of studies found that symptoms like fatigue and physical functioning were associated with outcomes like OS and PFS [135], [15]. Several of the models confirmed evidence that a better physical health before treatment may lead to better physical health after treatment, including a lower disease/symptom burden [42, 92]. This underlines that physical functioning and symptoms scales are important predictors for oncological outcomes.
HRQoL as outcome
The majority of the domains that have been included in prediction models as outcome are symptoms or a physical functioning scales; which is the same trend seen in the predictors. Physicians may favor symptoms over functioning domain as physicians may have more affinity or experience with these, for instance by registering CTCAE-graded symptoms. HRQoL outcomes were predicted by non-HRQoL factors treatment, comorbidities, age, gender, and previous malignancy. The noticeable absence of disease characteristics, like TNM staging or biomarkers, as predictors of HRQoL in the majority of publications was surprising, given their role in informing treatment decisions. However, from this observation it seems that treatment factors are favoured over disease characteristics as predictors of HRQoL as an outcome. An possible explanation is that treatment factors are more actionable variables for a clinician, and thus after prediction of HRQoL a treatment can be changed or adjusted to increase a patients chance on better HRQoL.
HRQoL as predictor and outcome
Several papers included HRQoL as both a predictor and outcome in one model and focused on predicting symptoms. For example, Dess et al. [40] found that baseline QoL score was predictive of erectile function in prostate cancer patients; Pfob et al. [39] showed that ‘satisfaction with breasts’ and ‘sexual wellbeing’ domains of the Breast-Q predicted satisfaction with breasts one year after surgery. The prediction of functioning domain scores with previous HRQoL scores seems to be a knowledge gap in the current literature. The prediction of HRQoL by HRQoL is still in its infancy as shown by the current models, but provide a promising avenue for future research. Combining HRQoL prediction models with previous research on association and clustering between HRQoL domains and distinct patient trajectories for HRQoL [136, 137], leads towards the possiblity to use prediction models to classify patients into risk categories. This classification is taking a first step towards tailored personalized treatment, which is the ultimate goal of personalized prediction.
Implications of the selected domains
The skewed focus on physical functioning and symptoms scales as predictors and outcomes reiterates the need to incorporate other HRQoL domains (emotional, role, cognitive functioning) in model development. This is important, as prediction with and of HRQoL could be a tool where patients and clinicians get insight in what can happen during treatment and disease under different treatment choices. Which HRQoL domains are important may differ from patient to patient and between cancer types. Therefore, limiting and skewing the models towards physical functioning and symptoms will limit the applicability in the clinic and intended population. The introduction of value based health care in oncology has shifted the focus from treating disease to treating the patient, which is exactly the shift the model development decisions must undergo.
Necessary steps for future HRQoL model development
As HRQoL in cancer prediction models is a rapidly evolving field, the results of this review point out several recommendations for future research to advance the field. First, the standard use of prediction model development and validation guidelines like TRIPOD and TRIPOD-AI is recommended to improve the model development standards and the poor reporting quality of models. Second, any HRQoL prediction model should at least have undergone internal validation to reduce overfitting and optimism in the model caused by the training data. Preferably, this should be followed by an external validation step if the models are intended for use in clinical practice. Third, other predictive methods using HRQoL data need to be tested, including dynamic models such as landmarking or joint modelling approaches that predict the next HRQoL time point based on the patient’s historical HRQoL while using all available data [138]. The dynamic aspect allows for better application and implementation of the model in the clinic, because it is one equation that can adjust its prediction at each new timepoint when new information becomes available [130].
Fourth, when designing the studies that underlie the next generation of HRQoL models, it is advised to power the sample size for AI based strategies in order to evaluate whether AI outperforms the classical methods. Most of the existing datasets of HRQoL do not reach sufficient sample size number, while some AI techniques require 200 events per candidate feature [139]. Of note, as HRQoL is a construct that measures the impact of the disease on every other domain of life, it is highly variable among different patient groups, cancer types, and even countries or regions. New studies in this field need to account for region/country differences, based on cultural and health care system differences, even when using validated PROMs [140]. To overcome both sample size as cultural differences issues, the internationally scattered HRQoL data needs to be brought together into one database for development of future prediction models.
Our final recommendation, is that all HRQoL domains should at least be included in the possible candidate predictors for the new generation models that predict any outcome for a cancer patient. This is needed to prevent the skew towards physical functioning and symptom scales, and thus possible include other domains (e.g. emotional or role) in the models that may be of more importance for cancer patients.
Strengths and limitations of the review
This review provides an overview of the emerging field of HRQoL prediction in oncology, which is important as HRQoL has become a central part of cancer treatment in recent years. This review lays the ground work for further development of HRQoL prediction models/field. A strength is the use of TRIPOD and PROBAST in the study, as recommended by the prediction modelling field [24].
One of the limitations of this study is the search terms, which were chosen to broadly capture prediction models, as proposed by Geersing et al. [21]. Due to the poor indexing of terms related to prediction and prediction modelling, it is possible to miss publications regarding HRQoL prediction modelling, even when using the Geersing filter. Specifically, our review may have missed relevant AI-based HRQoL prediction publication as no specific AI-terms were included in the used filter. Another limitation is the use of keywords surrounding the EORTC quality of life questionnaires in the search term. This may have biased this review towards finding a majority of studies that included these questionnaires. This may explain the relative lack of PROMIS and FACT questionnaires as it is clear from the HRQoL literature that they are used more frequently than was found in this review [141].
Future research
For future research, a review focussing solely on the AI HRQoL prediction models is needed as the current review includes mainly explorative models in this field. A review on machine-learning and PRO’s in oncology is currently being conducted by Krepper et al. [142]. Further research should also explore the impact of the dissemination through scientific literature, especially due to the varying nature of the publishing quality of the articles itself as well as the impact of the journal it was published in. Most of the articles (76%) were published in a first quartile ranked journal (SCImago journal ranking (SJR)), which ranges from first quartile to the fourth quartile based on a weighted average of how many times the previous three years (e.g. 2020–2022) of publications are cited in the year after (2023) (SCImago). Reporting quality of the model thus did not impact publication in high ranking journals.
All publications in this review described the development and/or validation of prediction models but did not address their further implementation or practical applications. The next step should be to evaluate the clinical use of prediction models. This will help accelerate understanding on which models are both feasible and desired by clinicians and patients.
Conclusion
To conclude, prediction modelling of and with HRQoL in oncology is still in its infancy. However, this review underscores the potential of using HRQoL as a tool for predicting disease outcomes and as a valuable resource for informing patients about the potential consequences of their disease and treatment. To enhance the transparency of reporting, the authors would strongly advocate adherence to the TRIPOD (-AI) checklist for development and validation in the design and reporting of any HRQoL prediction model study [23, 132].
Acknowledgements
The authors are grateful for the support of the Netherlands Cancer Institute librarians for the support with retrieving all publications.
Declarations
Competing interests
TvdH: None to declare, KdL: None to declare, SvdM: None to declare, NH: None to declare, BH: None to declare, LvdP: None to declare, BdR: None to declare.
Ethical approval
No approval needed for this literature review.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Binnen de bundel kunt u gebruik maken van boeken, tijdschriften, e-learnings, web-tv's en uitlegvideo's. BSL Podotherapeut Totaal is overal toegankelijk; via uw PC, tablet of smartphone.
Exploring the role of health-related quality of life measures in predictive modelling for oncology: a systematic review
Auteurs
T. G. W. van der Heijden K. M. de Ligt N. J. Hubel S. van der Mierden B. Holzner L. V. van de Poll-Franse B. H. de Rooij the EORTC Quality of Life Group
