Background
According to the World Health Organization, cancer is one of the leading causes of morbidity and mortality worldwide, with lung cancer in the top five of cancers and the leading cause of cancer mortality with 1,6 million deaths in 2012 [
1]. Roughly 80–85% of lung cancers are non-small cell lung cancer (NSCLC) [
2]. Staging in patients is based on the Tumor Node Metastasis (TNM) classification, which is shown to be an important predictor of survival [
3,
4]. Incurable patients with initial or recurrent metastatic NSCLC (stages IIIB and IV) have a short life expectancy, with 1-, 2- and 3-years survival ranging between 22 and 47%, 8–26% and 4–17%, respectively [
5].
The current treatment of incurable NSCLC patients consists of systemic chemotherapy (CT), radiotherapy (RT), therapies targeting oncodrivers (e.g., epidermal growth factor receptor tyrosine kinase inhibitors, EGFR-TKI) or the immune system (immunotherapies), in addition to best supportive care (e.g., pain relief). In specific patient groups, mainly palliative surgery (for spinal metastases) and RT (for brain metastases) are advised. Palliative treatments aim to preserve or improve quality of life, lengthen life or decrease disease burden. Palliation can target the tumor tissue itself or symptoms, such as pain, diarrhea, obstipation, anxiety or depression. Individually tailored palliative cancer treatment is essential to ensure that patients receive the treatment that optimally matches their values and preferences, avoiding under- or overtreatment, and optimally utilizing available healthcare resources. However, this is a challenge due to the heterogeneity of the patient population, the multiple treatment options, and the marginal expected treatment benefits. Therefore, decision making in the palliative phase can be complex, as there is a delicate balance between benefits (e.g., symptom relief, life lengthening) and harms of treatments (e.g., side effects, loss of quality of life), as well as the costs of treatment. Decision support systems (DSS) could assist physicians in formulating an evidence-based treatment advice. DSS (e.g., prediction models, nomograms or decision trees) are based on statistical models in order to predict outcomes, such as overall survival (OS) (with our without treatment), toxicity and cost-effectiveness. They are based on patient and tumor characteristics, and preferably compare various treatment options. Research has shown that such clinical prediction models in end-of-life care are valued by physicians, because they enhance prognostic confidence and improve communication with patients, although they can also cause emotional distress in patients and raise prognostic overconfidence despite uncertainty in palliative care [
6].
In patients with incurable NSCLC, some overviews have been published that summarize DSS in NSCLC patients. For instance, Mahar et al. have performed a systematic literature search from 1996 until 2015, identifying a total of 32 tools for all stages of lung cancer [
7]. They described that the majority of the prediction models focus on NSCLC patients with metastatic disease, which can be explained by a larger need for DSS in this specific clinical population [
7]. However, they did not use an extensive literature search with a large variety of MeSH headings, and thus, might have missed DSS for this subgroup. Other reviews have described DSS specifically developed either for patients with spinal metastases [
8] or brain metastases, largely consisting of incurable NSCLC patients [
9‐
12]. However, none of the earlier studies focused on the available DSS for the entire incurable NSCLC population, having short survival times due to rapidly progressive disease, whilst on the other hand there are rapid developments of new treatment options. Tools that aid clinical decision-making in this complex subgroup are urgently needed to help oncologists navigate the ever-growing maze of treatment options.
We conducted an extensive systematic literature search in order to summarize the available DSS for incurable patients with (initial or recurrent) metastatic NSCLC (stages IIIB and IV). We will give an overview of the development studies and the included predictors, as well as the levels of validation and calibration, and the model performances. Furthermore, we add concluding remarks about the user friendliness and ease of access of the identified DSS in clinical practice, and give direction to future research in this rapidly evolving field.
Discussion
Decision support systems (DSS) aid clinical decision-making by comparing various treatment options, and by predicting harms and benefits based on patient and tumor characteristics. This systematic review provides a comprehensive overview of DSS developed and/or validated for incurable patients with (initial or recurrent) metastatic NSCLC (stages IIIB and IV). In total, 39 DSS have been identified, of which 17 had been externally validated. Each DSS is described, and an overview is given of their discrimination, calibration and user friendliness. These DSS estimate OS and/or PFS, and are based on patient and treatment characteristics, sociodemographic, lifestyle and physical factors, serum markers, and to a lesser extent genetic markers. Regardless of the relatively large amount of existing DSS, there is room for improvement in the tools for clinical decision-making.
Less than half of the currently available DSS have been externally validated in a broader setting, and most validations have also been performed in relatively old datasets. Validated tools also showed poor model performances. Another shortcoming of the currently available DSS is that they only estimate OS or PFS, but do not incorporate other outcomes of societal relevance, such as toxicity or cost-effectiveness. In line with previous reviews, none of the DSS weigh the risks and benefits of treatments [
7,
9‐
12]. This makes decision-making difficult, as not all facets are discussed. It should be kept in mind that DSS only aim to facilitate the decision-making process. Physician can use information obtained from DSS to derive a treatment advice, or to inform patients during consultations. The information obtained from DSS can help patients develop informed preferences, which are the basis for shared decision-making. It is therefore important for DSS to at least provide information of both the benefits (in terms of survival) and harms (in terms of side-effects) of treatment.
Also, most tools are developed to give a rough estimate of survival, either in the entire incurable NSCLC population, for systemic therapy, targeted therapy, mixed treatments or specifically for patients with brain or spinal metastases. There is no DSS that gives an overview of all treatments relevant to consider in the incurable NSCLC population (or a specific subgroup), or that offers clear cut-off points for when it is worthwhile to provide intensive treatment or best supportive care. In the meantime, studies often lack good definition of the control conditions, which are described as a ‘more conservative approach’ or ‘best supportive care’. Even though clinical guidelines also describe all available treatment options, they do not present overviews that enable individualized decision making. Some currently identified DSS are incorporated in existing guidelines, although these tools’ performance is mediocre. For instance, the RPA was extensively examined in multiple studies and is incorporated in the ESMO guidelines [
56] and the Dutch Oncoline guidelines for brain metastases [
55], but its discriminative ability is not at all strong. The revised Tokuhashi score is only mentioned in the Dutch national guidelines for spinal metastases [
62], although the accuracy of this DSS is not consistently good in all studies [
58,
59]. For more personalized clinical decision-making, guidelines would ideally incorporate available DSS based on recent clinical evidence with good discriminatory ability and calibration that compare multiple treatment options, and present multiple outcomes (e.g., benefits, harms and cost-effectiveness).
Within this review, we found that DSS that outperformed others (e.g., Di Maio score with AUC = 0.926) or that have a user friendly lay-out (e.g., Barnholtz-Sloan nomogram) are not validated in broader settings, and have not been tested extensively. An explanation for the lack of optimal tools could be that the current process of development, validation and updating of DSS is too time-consuming. Therefore, DSS are expected to be outdated by the time that extensive validations can be performed due to the rapid developments in lung cancer care. Other methods, such as rapid learning techniques [
68,
69] and other sophisticated algorithms might be developed for more continuous updating and validating procedures. Another explanation for the relatively poor model performances might also be that survival in this heterogeneous group of patients cannot be estimated with high accuracy. Even though they aim to personalize decision-making, DSS are based on statistical models that by definition make use of probabilities. These models generally describe the association between an outcome and a very limited set of potential predictors only. By adding a broader range of predictors from large longitudinal databases to build and validate models, perhaps the biological complexity and heterogeneity can be better reflected in the resulting outcome predictions. The use of biomarkers might lead to higher accuracy than the more general predictors such as age and tobacco use.
In the last decennia, the concept of personalized medicine has taken a more central position in metastasized cancer care. Therefore, future DSS should take into account specific biological markers and genes, such as
EGFR and
ALK. The complexity of gene mutations, translocations and rearrangements can explain why some treatments are effective, while others induce little response in patients. Some systematic reviews have summarized the currently known and relevant NSCLC genetic markers (e.g.,
EGFR, EML4/ALK mutations), and other markers for which there is insufficient evidence for use in clinical decision-making (e.g.,
K-RAS,
ERCC1
, BRCA, Beta tubulin III, RRM1, TP-53 mutations) [
70‐
72]. Genetic markers will become increasingly important in the future in order to distinguish between responders and non-responders. The same is true for immunotherapeutic approaches, as for example, the NCCN guidelines recommend immune checkpoint inhibitors for incurable NSCLC patients, based on performance status and treatment responses [
73]. These guidelines give some insights and flow diagrams about the application of two new immunotherapeutic agents, nivolumab and pembrolizumab that target the programmed cell death protein 1 (PD-1) pathway. However, important insights are still lacking to determine for which patients immunological treatments are (most) effective, especially considering that the targeted treatments are costly and some induce severe side effects.
The current systematic review aimed to shed light on which tools are available, and which gaps remain to be filled in future research. In general, the available DSS are of limited value to daily clinical practice because they used relatively old clinical data (before 2000), focused more on advantages than disadvantages of merely one or two treatment options, and still lack available user friendly applications. By collaborating with national databases, a continuous updating procedure could be incorporated as well. Recently some research groups have made new web-based prediction tools that shed light on both the advantages and disadvantages of treatments. One example is the Predict Cancer website [
74], where an application is presented including DSS for lung, rectum, head and neck cancer and brain metastases. This application aims to support oncologists with the estimation of expected survival rates, side effects of treatments, cost-effectiveness of a treatment plan, and other important parameters. As another example, the ASCO, ESMO, NCCN and Institute for Clinical and Economic Review (ICER) have presented frameworks, in which not only benefits but also the cost-effectiveness and toxicity of several anti-cancer drugs were quantified [
75,
76]. Furthermore, Warner et al. have created a novel and promising rapid learning system for various cancer types (including lung cancer) that automatically calculates and displays mutation-specific survival rates from electronic health record data (69). These approaches could be useful for future DSS too.
Strengths of the current systematic review are the extensive literature search performed both in online databases and with reference tracking. We have chosen a specific population in which decision support can be of great value because of the short survival time, rapid treatment-related developments, and complex decisions about when it still is in the patients’ best interest to provide invasive treatment, or when the transition needs to be made to provide only symptom relief. Nevertheless, some limitations have to be taken into account as well. Terminology to describe DSS varies greatly throughout the field, and this could have hampered the ability to find all existing DSS even with an extensive search strategy. Also, we have not used the complete CHARMS checklist [
15] to assess methodological quality of the included DSS, as many items were not reported. Instead we created an abbreviated list of items that covers the main aspects of model performance (shown in Table
4).
In the current study, 39 DSS have been identified for incurable metastatic NSCLC patients. Previously, Mahar et al. have performed a systematic literature search, and found 32 tools both for small cell lung cancer (
N = 7) and NSCLC (
N = 25) [
7]. Of the tools developed for NSCLC patients, there was one tool for all tumor stages, eight tools for stages I-III, and 16 for advanced/incurable disease [
7]. Mahar et al. found 16 tools for our target population, of which three were excluded in our current search: two were based on older clinical data (before 2000), and one only reported hazard ratios and no formulas for individual probabilities. Furthermore, three of their included tools were published together in one paper, and were summarized as one tool in our current study [
77]. To conclude, our systematic literature search identified 28 additional studies, and added extensive information about the studies that have externally validated these DSS, the user friendliness and the application methods of the tool in clinical practice.
Conclusions
Our study adds to the current knowledge in the field of DSS in incurable NSCLC, as conclusions are drawn about the extent and quality of validation (i.e., Reilly levels of validation), extent of calibration (i.e., Van Calster levels of calibration) and user friendliness (i.e., routine collection, ease of use, online tools). In addition, an overview is given of the used predictors, grouped into domains. Not only does our review increases knowledge of existing DSS, but we also indicate areas for improvement. Overall, we can conclude that multiple DSS have been developed for incurable patients with (initial or recurrent) metastatic NSCLC (stages IIIB and IV), but most of them have used relatively old clinical data, focused on benefits rather than harms in terms of toxicity and risks, did not compare various treatment options, control treatments (conservative treatment, best supportive care, usual care) were often poorly or not described, have not been (externally) validated or still lack available user friendly applications (i.e., scoring tables, online calculators, mobile applications). Also, various predictors in the domains of serum markers, tumor and treatment characteristics have been included, but apart from EGFR in one DSS, biological markers for targeted and immunotherapies are still lacking. Other methods might be available for future DSS designs in order to incorporate all relevant individual characteristics efficiently, while taking into account the needs of oncologists in their daily practice. Preferably, future DSS provide oncologists an efficient method to stay up-to-date with the rapid innovations in lung cancer care.