A Two-Step Framework for Validating Causal Effect Estimates.

BACKGROUND: Comparing causal effect estimates obtained using observational data to those obtained from the gold standard (i.e., randomized controlled trials [RCTs]) helps assess the validity of these estimates. However, comparisons are challenging due to differences between observational data and RCT generated data. The unknown treatment assignment mechanism in the observational data and varying sampling mechanisms between the RCT and the observational data can lead to confounding and sampling bias, respectively. AIMS: The objective of this study is to propose a two-step framework to validate causal effect estimates obtained from observational data by adjusting for both mechanisms. MATERIALS AND METHODS: An estimator of causal effects related to the two mechanisms is constructed. A two-step framework for comparing causal effect estimates is derived from the estimator. An R package RCTrep is developed to implement the framework in practice. RESULTS: A simulation study is conducted to show that using our framework observational data can produce causal effect estimates similar to those of an RCT. A real-world application of the framework to validate treatment effects of adjuvant chemotherapy obtained from registry data is demonstrated. CONCLUSION: This  study constructs a framework for comparing causal effect estimates between observational data and RCT data, facilitating the assessment of the validity of causal effect estimates obtained from observational data.

Value of Real-World Evidence for Treatment Selection: A Case Study in Colon Cancer.

PURPOSE: Real-world evidence (RWE)-derived from analysis of real-world data (RWD)-has the potential to guide personalized treatment decisions. However, because of potential confounding, generating valid RWE is challenging. This study demonstrates how to responsibly generate RWE for treatment decisions. We validate our approach by demonstrating that we can uncover an existing adjuvant chemotherapy (ACT) guideline for stage II and III colon cancer (CC)-which came about using both data from randomized controlled trials and expert consensus-solely using RWD. METHODS: Data from the population-based Netherlands Cancer Registry from a total of 27,056 patients with stage II and III CC who underwent curative surgery were analyzed to estimate the overall survival (OS) benefit of ACT. Focusing on 5-year OS, the benefit of ACT was estimated for each patient using G-computation methods by adjusting for patient and tumor characteristics and estimated propensity score. Subsequently, on the basis of these estimates, an ACT decision tree was constructed. RESULTS: The constructed decision tree corresponds to the current Dutch guideline: patients with stage III or stage II with T stage 4 should receive surgery and ACT, whereas patients with stage II with T stage 3 should only receive surgery. Interestingly, we do not find sufficient RWE to conclude against ACT for stage II with T stage 4 and microsatellite instability-high (MSI-H), a recent addition to the current guideline. CONCLUSION: RWE, if used carefully, can provide a valuable addition to our construction of evidence on clinical decision making and therefore ultimately affect treatment guidelines. Next to validating the ACT decisions advised in the current Dutch guideline, this paper suggests additional attention should be paid to MSI-H in future iterations of the guideline.

Head and neck cancers survival in Europe, Taiwan, and Japan: results from RARECAREnet Asia based on a privacy-preserving federated infrastructure.

Background: The head and neck cancers (HNCs) incidence differs between Europe and East Asia. Our objective was to determine whether survival of HNC also differs between European and Asian countries. Methods: We used population-based cancer registry data to calculate 5-year relative survival (RS) for the oral cavity, hypopharynx, larynx, nasal cavity, and major salivary gland in Europe, Taiwan, and Japan. We modeled RS with a generalized linear model adjusting for time since diagnosis, sex, age, subsite, and histological grouping. Analyses were performed using federated learning, which enables analyses without sharing sensitive data. Findings: Five-year RS for HNC varied between geographical areas. For each HNC site, Europe had a lower RS than both Japan and Taiwan. HNC subsites and histologies distribution and survival differed between the three areas. Differences between Europe and both Asian countries persisted even after adjustments for all HNC sites but nasal cavity and paranasal sinuses, when comparing Europe and Taiwan. Interpretation: Survival differences can be attributed to different factors including different period of diagnosis, more advanced stage at diagnosis, or different availability/access of treatment. Cancer registries did not have stage and treatment information to further explore the reasons of the observed survival differences. Our analyses have confirmed federated learning as a feasible approach for data analyses that addresses the challenges of data sharing and urge for further collaborative studies including relevant prognostic factors.

Estimating Treatment Effect of Adjuvant Chemotherapy in Elderly Patients With Stage III Colon Cancer Using Bayesian Networks.

PURPOSE: While adjuvant therapy with capecitabine and oxaliplatin (CAPOX) has been proven to be effective in stage III colon cancer, capecitabine monotherapy (CapMono) might be equally effective in elderly patients. Unfortunately, the elderly are under-represented in clinical trials and patients included may not be representative of the routine care population. Observational data might alleviate this problem but is sensitive to biases such as confounding by indication. Here, we build causal models using Bayesian Networks (BNs), identify confounders, and estimate the effect of adjuvant chemotherapy using survival analyses. METHODS: Patients 70 years and older were selected from the Netherlands Cancer Registry (N = 982). We developed several BNs using constraint-based, score-based, and hybrid algorithms while precluding noncausal relations. In addition, we created models using a limited set of recurrence and survival nodes. Potential confounders were identified through the resulting graphs. Several Cox models were fitted correcting for confounders and for propensity scores. RESULTS: When comparing adjuvant treatment with surgery only, pathological lymph node classification, physical status, and age were identified as potential confounders. Adjuvant treatment was significantly associated with survival in all Cox models, with hazard ratios between 0.39 and 0.45; CIs overlapped. BNs investigating CAPOX versus CapMono did not find any association between the treatment choice and survival and thus no confounders. Analyses using Cox models did not identify significant association either. CONCLUSION: We were able to successfully leverage BN structure learning algorithms in conjunction with clinical knowledge to create causal models. While confounders differed depending on the algorithm and included nodes, results were not contradictory. We found a strong effect of adjuvant therapy on survival in our cohort. Additional oxaliplatin did not have a marked effect and should be avoided in elderly patients.

Comparing quality of breast cancer care in the Netherlands and Norway by federated propensity score analytics.

PURPOSE: The aim of the study was to benchmark and compare breast cancer care quality indicators (QIs) between Norway and the Netherlands using federated analytics preventing transfer of patient-level data. METHODS: Breast cancer patients (2017-2018) were retrieved from the Netherlands Cancer Registry and the Cancer Registry of Norway. Five European Society of Breast Cancer Specialists (EUSOMA) QIs were assessed: two on magnetic resonance imaging (MRI), two on surgical approaches, and one on postoperative radiotherapy. The QI outcomes were calculated using 'Vantage 6' federated Propensity Score Stratification (PSS). Likelihood of receiving a treatment was expressed in odds ratios (OR). RESULTS: In total, 39,163 patients were included (32,786 from the Netherlands and 6377 from Norway). PSS scores were comparable to the crude outcomes of the QIs. The Netherlands scored higher on the QI 'proportions of patients preoperatively examined with breast MRI' [37% vs.17.5%; OR 2.8 (95% CI 2.7-2.9)], the 'proportions of patients receiving primary systemic therapy examined with breast MRI' [83.3% vs. 70.8%; OR 2.3 (95% CI 1.3-3.3)], and 'proportion of patients receiving a single breast operation' [95.2% vs. 91.5%; OR 1.8 (95% CI 1.4-2.2)]. Country scores for 'immediate breast reconstruction' and 'postoperative radiotherapy after breast-conserving surgery' were comparable. The EUSOMA standard was achieved in both countries for 4/5 indicators. CONCLUSION: Both countries achieved high scores on the QIs. Differences were observed in the use of MRI and proportion of patients receiving single surgery. The federated approach supports future possibilities on benchmark QIs without transfer of privacy-sensitive data.

Identifying Confounders Using Bayesian Networks and Estimating Treatment Effect in Prostate Cancer With Observational Data.

PURPOSE: Randomized controlled trials are considered the golden standard for estimating treatment effect but are costly to perform and not always possible. Observational data, although readily available, is sensitive to biases such as confounding by indication. Structure learning algorithms for Bayesian Networks (BNs) can be used to discover the underlying model from data. This enables identification of confounders through graph analysis, although the model might contain noncausal edges. We propose using a blacklist to aid structure learning in finding causal relationships. This is illustrated by an analysis into the effect of active treatment (v observation) in localized prostate cancer. METHODS: In total, 4,121 prostate cancer records were obtained from the Netherlands Cancer Registry. Subsequently, we developed a (causal) BN using structure learning while precluding noncausal relations. Additionally, we created several Cox proportional hazards models, each correcting for a different set of potential confounders (including propensity scores). Model predictions for overall survival were compared with expected survival on the basis of the general population using data from Statistics Netherlands (Centraal Bureau voor de Statistiek). RESULTS: Structure learning precluding noncausal relations resulted in a causal graph but did not identify significant edges toward treatment; they were added manually. Graph analysis identified year of diagnosis and age as confounders. The BN predicted a treatment effect of 1 percentage point at 10 years. Chi-squared analysis found significant associations between year of diagnosis, age, stage, and treatment. Propensity score correction was successful. Adjusted Cox models predicted significant treatment effect around 3 percentage points at 10 years. CONCLUSION: A blacklist in conjunction with structure learning can result in a causal BN that can be used for confounder identification. Treatment effect found here is close to the 5 percentage point found in the literature.

An Improved Infrastructure for Privacy-Preserving Analysis of Patient Data.

Incorporating healthcare data from different sources is crucial for a better understanding of patient (sub)populations. However, data centralization raises concerns about data privacy and governance. In this work, we present an improved infrastructure that allows privacy-preserving analysis of patient data: vantage6 v3. For this new version, we describe its architecture and upgraded functionality, which allows algorithms running at each party to communicate with one another through a virtual private network (while still being isolated from the public internet to reduce the risk of data leakage). This allows the execution of different types of algorithms (e.g., multi-party computation) that were practically infeasible before, as showcased by the included examples. The (continuous) development of this type of infrastructure is fundamental to meet the current and future demands of healthcare research with a strong emphasis on preserving the privacy of sensitive patient data.

A Narrative Review on the Collection and Use of Electronic Patient-Reported Outcomes in Cancer Survivorship Care with Emphasis on Symptom Monitoring.

Electronic patient-reported outcome (ePRO) applications promise great added value for improving symptom management and health-related quality of life. The aim of this narrative review is to describe the collection and use of ePROs for cancer survivorship care, with an emphasis on ePRO-symptom monitoring. It offers many different perspectives from research settings, while current implementation in routine care is ongoing. ePRO collection optimizes survivorship care by providing insight into the patients' well-being and prioritizing their unmet needs during the whole trajectory from diagnosis to end-of-life. ePRO-symptom monitoring can contribute to timely health risk detection and subsequently allow earlier intervention. Detection is optimized by automatically generated alerts that vary from simple to complex and multilayered. Using ePRO-symptoms during in-hospital consultation enhances the patients' conversation with the health care provider before making informed decisions about treatments, other interventions, or self-management. ePRO(-symptoms) entail specific implementation issues and complementary ethics considerations. The latter is due to privacy concerns, digital divide, and scarcity of adequately representative data for particular groups of patients.

Exploring Cancer Survivor Needs and Preferences for Communicating Personalized Cancer Statistics From Registry Data: Qualitative Multimethod Study.

BACKGROUND: Disclosure of cancer statistics (eg, survival or incidence rates) based on a representative group of patients can help increase cancer survivors' understanding of their own diagnostic and prognostic situation, and care planning. More recently, there has been an increasing interest in the use of cancer registry data for disclosing and communicating personalized cancer statistics (tailored toward personal and clinical characteristics) to cancer survivors and relatives. OBJECTIVE: The aim of this study was to explore breast cancer (BCa) and prostate cancer (PCa) survivor needs and preferences for disclosing (what) and presenting (how) personalized statistics from a large Dutch population-based data set, the Netherlands Cancer Registry (NCR). METHODS: To elicit survivor needs and preferences for communicating personalized NCR statistics, we created different (non)interactive tools visualizing hypothetical scenarios and adopted a qualitative multimethod study design. We first conducted 2 focus groups (study 1; n=13) for collecting group data on BCa and PCa survivor needs and preferences, using noninteractive sketches of what a tool for communicating personalized statistics might look like. Based on these insights, we designed a revised interactive tool, which was used to further explore the needs and preferences of another group of cancer survivors during individual think-aloud observations and semistructured interviews (study 2; n=11). All sessions were audio-recorded, transcribed verbatim, analyzed using thematic (focus groups) and content analysis (think-aloud observations), and reported in compliance with qualitative research reporting criteria. RESULTS: In both studies, cancer survivors expressed the need to receive personalized statistics from a representative source, with especially a need for survival and conditional survival rates (ie, survival rate for those who have already survived for a certain period). Personalized statistics adjusted toward personal and clinical factors were deemed more relevant and useful to know than generic or average-based statistics. Participants also needed support for correctly interpreting the personalized statistics and putting them into perspective, for instance by adding contextual or comparative information. Furthermore, while thinking aloud, participants experienced a mix of positive (sense of hope) and negative emotions (feelings of distress) while viewing the personalized survival data. Overall, participants preferred simplicity and conciseness, and the ability to tailor the type of visualization and amount of (detailed) statistical information. CONCLUSIONS: The majority of our sample of cancer survivors wanted to receive personalized statistics from the NCR. Given the variation in patient needs and preferences for presenting personalized statistics, designers of similar information tools may consider potential tailoring strategies on multiple levels, as well as effective ways for providing supporting information to make sure that the personalized statistics are properly understood. This is encouraging for cancer registries to address this unmet need, but also for those who are developing or implementing personalized data-driven information tools for patients and relatives.

Explainable machine learning can outperform Cox regression predictions and provide insights in breast cancer survival.

Cox Proportional Hazards (CPH) analysis is the standard for survival analysis in oncology. Recently, several machine learning (ML) techniques have been adapted for this task. Although they have shown to yield results at least as good as classical methods, they are often disregarded because of their lack of transparency and little to no explainability, which are key for their adoption in clinical settings. In this paper, we used data from the Netherlands Cancer Registry of 36,658 non-metastatic breast cancer patients to compare the performance of CPH with ML techniques (Random Survival Forests, Survival Support Vector Machines, and Extreme Gradient Boosting [XGB]) in predicting survival using the [Formula: see text]-index. We demonstrated that in our dataset, ML-based models can perform at least as good as the classical CPH regression ([Formula: see text]-index [Formula: see text]), and in the case of XGB even better ([Formula: see text]-index [Formula: see text]). Furthermore, we used Shapley Additive Explanation (SHAP) values to explain the models' predictions. We concluded that the difference in performance can be attributed to XGB's ability to model nonlinearities and complex interactions. We also investigated the impact of specific features on the models' predictions as well as their corresponding insights. Lastly, we showed that explainable ML can generate explicit knowledge of how models make their predictions, which is crucial in increasing the trust and adoption of innovative ML techniques in oncology and healthcare overall.

Heterogeneity in Quality of Life of Long-Term Colon Cancer Survivors: A Latent Class Analysis of the Population-Based PROFILES Registry.

BACKGROUND: Long-term colon cancer survivors present heterogeneous health-related quality of life (HRQOL) outcomes. We determined unobserved subgroups (classes) of survivors with similar HRQOL patterns and investigated their stability over time and the association of clinical covariates with these classes. MATERIALS AND METHODS: Data from the population-based PROFILES registry were used. Included were survivors with nonmetastatic (TNM stage I-III) colon cancer (n = 1,489). HRQOL was assessed with the Dutch translation of the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire C30 version 3.0. Based on survivors' HRQOL, latent class analysis (LCA) was used to identify unobserved classes of survivors. Moreover, latent transition analysis (LTA) was used to investigate changes in class membership over time. Furthermore, the effect of covariates on class membership was assessed using multinomial logistic regression. RESULTS: LCA identified five classes at baseline: class 1, excellent HRQOL (n = 555, 37.3%); class 2, good HRQOL with prevalence of insomnia (n = 464, 31.2%); class 3, moderate HRQOL with prevalence of fatigue (n = 213, 14.3%); class 4, good HRQOL with physical limitations (n = 134, 9.0%); and class 5, poor HRQOL (n = 123, 8.3%). All classes were stable with high self-transition probabilities. Longer time since the diagnosis, no comorbid conditions, and male sex were associated with class 1, whereas older age was associated with class 4. Clinical covariates were not associated with class membership. CONCLUSION: The identified classes are characterized by distinct patterns of HRQOL and can support patient-centered care. LCA and LTA are powerful tools for investigating HRQOL in cancer survivors. IMPLICATIONS FOR PRACTICE: Long-term colon cancer survivors show great heterogeneity in their health-related quality of life. This study identified five distinct clusters of survivors with similar patterns of health-related quality of life and showed that these clusters remain stable over time. It was also shown that these clusters do not significantly differ in tumor characteristics or received treatment. Cluster membership of long-term survivors can be identified by sociodemographic characteristics but is not predetermined by diagnosis and treatment.

Prognostic factors analysis for oral cavity cancer survival in the Netherlands and Taiwan using a privacy-preserving federated infrastructure.

The difference in incidence of oral cavity cancer (OCC) between Taiwan and the Netherlands is striking. Different risk factors and treatment expertise may result in survival differences between the two countries. However due to regulatory restrictions, patient-level analyses of combined data from the Netherlands and Taiwan are infeasible. We implemented a software infrastructure for federated analyses on data from multiple organisations. We included 41,633‬ patients with single-tumour OCC between 2004 and 2016, undergoing surgery, from the Taiwan Cancer Registry and Netherlands Cancer Registry. Federated Cox Proportional Hazard was used to analyse associations between patient and tumour characteristics, country, treatment and hospital volume with survival. Five factors showed differential effects on survival of OCC patients in the Netherlands and Taiwan: age at diagnosis, stage, grade, treatment and hospital volume. The risk of death for OCC patients younger than 60 years, with advanced stage, higher grade or receiving adjuvant therapy after surgery was lower in the Netherlands than in Taiwan; but patients older than 70 years, with early stage, lower grade and receiving surgery alone in the Netherlands were at higher risk of death than those in Taiwan. The mortality risk of OCC in Taiwanese patients treated in hospitals with higher hospital volume (≥ 50 surgeries per year) was lower than in Dutch patients. We conducted analyses without exchanging patient-level information, overcoming barriers for sharing privacy sensitive information. The outcomes of patients treated in the Netherlands and Taiwan were slightly different after controlling for other prognostic factors.

Machine Learning Explainability in Breast Cancer Survival.

Machine Learning (ML) can improve the diagnosis, treatment decisions, and understanding of cancer. However, the low explainability of how "black box" ML methods produce their output hinders their clinical adoption. In this paper, we used data from the Netherlands Cancer Registry to generate a ML-based model to predict 10-year overall survival of breast cancer patients. Then, we used Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) to interpret the model's predictions. We found that, overall, LIME and SHAP tend to be consistent when explaining the contribution of different features. Nevertheless, the feature ranges where they have a mismatch can also be of interest, since they can help us identifying "turning points" where features go from favoring survived to favoring deceased (or vice versa). Explainability techniques can pave the way for better acceptance of ML techniques. However, their evaluation and translation to real-life scenarios need to be researched further.

Predicting Lung Cancer Survival Using Probabilistic Reclassification of TNM Editions With a Bayesian Network.

PURPOSE: The TNM classification system is used for prognosis, treatment, and research. Regular updates potentially break backward compatibility. Reclassification is not always possible, is labor intensive, or requires additional data. We developed a Bayesian network (BN) for reclassifying the 5th, 6th, and 7th editions of the TNM and predicting survival for non-small-cell lung cancer (NSCLC) without training data with known classifications in multiple editions. METHODS: Data were obtained from the Netherlands Cancer Registry (n = 146,084). A BN was designed with nodes for TNM edition and survival, and a group of nodes was designed for all TNM editions, with a group for edition 7 only. Before learning conditional probabilities, priors for relations between the groups were manually specified after analysis of changes between editions. For performance evaluation only, part of the 7th edition test data were manually reclassified. Performance was evaluated using sensitivity, specificity, and accuracy. Two-year survival was evaluated with the receiver operating characteristic area under the curve (AUC), and model calibration was visualized. RESULTS: Manual reclassification of 7th to 6th edition stage group as ground truth for testing was impossible in 5.6% of the patients. Predicting 6th edition stage grouping using 7th edition data and vice versa resulted in average accuracies, sensitivities, and specificities between 0.85 and 0.99. The AUC for 2-year survival was 0.81. CONCLUSION: We have successfully created a BN for reclassifying TNM stage grouping across TNM editions and predicting survival in NSCLC without knowing the true TNM classification in various editions in the training set. We suggest binary prediction of survival is less relevant than predicted probability and model calibration. For research, probabilities can be used for weighted reclassification.

VANTAGE6: an open source priVAcy preserviNg federaTed leArninG infrastructurE for Secure Insight eXchange.

Answering many of the research questions in the field of cancer informatics requires incorporating and centralizing data that are hosted by different parties. Federated Learning (FL) has emerged as a new approach in which a global model can be generated without disclosing private patient data by keeping them at their original location. Flexible, user-friendly, and robust infrastructures are crucial for bringing FL solutions to the day-to-day work of the cancer epidemiologist. In this paper, we present an open source priVAcy preserviNg federaTed leArninG infrastructurE for Secure Insight eXchange, VANTAGE6. We provide a detailed description of its conceptual design, modular architecture, and components. We also show a few examples where VANTAGE6 has been successfully used in research on observational cancer data. Developing and deploying technology to support federated analyses - such as VANTAGE6 - will pave the way for the adoption and mainstream practice of this new approach for analyzing decentralized data.

Communicative aspects of decision aids for localized prostate cancer treatment - A systematic review.

CONTEXT: Despite increasing interest in the development and use of decision aids (DAs) for patients with localized prostate cancer (LPC), little attention has been paid to communicative aspects (CAs) of such tools. OBJECTIVE: To identify DAs for LPC treatment, and review these tools for various CAs. MATERIALS AND METHODS: DAs were identified through both published literature (MEDLINE, Embase, CINAHL, CENTRAL, and PsycINFO; 1990-2018) and online sources, in compliance with the Preferred Reporting Items for Systematic Review and Meta-Analyses guidelines. Identified DAs were reviewed for the International Patient Decision Aid Standards criteria, and analyzed on CAs, including information presentation, personalization, interaction, information control, accessibility, suitability, and source of information. Nineteen DAs were identified. RESULTS: International Patient Decision Aid Standards scores varied greatly among DAs. Crucially, substantial variations in use of CAs by DAs were identified: (1) few DAs used visual aids to communicate statistical information, (2) none were personalized in terms of outcome probabilities or mode of communication, (3) a minority used interactive methods to elicit patients' values and preferences, (4) most included biased cross tables to compare treatment options, and (5) issues were observed in suitability and accessibility that could hinder implementation in clinical practice. CONCLUSIONS: Our review suggests that DAs for LPC treatment could be further improved by adding CAs such as personalized outcome predictions and interaction methods to the DAs. Clinicians who are using or developing such tools might therefore consider these CAs in order to enhance patient participation in treatment decision-making.