Skip to main content

Table 1 Summary of included 13 papers (See Supplementary File 1 for the completed extraction table)

From: A scoping review of causal methods enabling predictions under hypothetical interventions

Title Intervention Scenario Clinical topic area Types of outcomes Stated assumptions Reported limitations Code availability
Candido dos Reis, F. J. et al. (2017) An updated PREDICT breast cancer prognostication and treatment benefit prediction model with independent validation, Breast Cancer Research, 19(1), 58 Single intervention, Discrete choice, Average effect Breast cancer Survival Generalisability of effect from clinical trial. Prediction of non-breast cancer deaths was excellent in the model development data set but could under-predict or over-predict in the validation data sets. Stata code are available from the author on request.
Brunner, F. J. et al. (2019) Application of non-HDL cholesterol for population-based cardio-vascular risk stratification: results from the Multinational Cardiovascular Risk Consortium., The Lancet 394.10215: 2173-2183. Single intervention, Discrete choice, Average effect CVD Binary The therapeutic benefit of lipid-lowering intervention investigated in the study is based on a hypothetical model that assumes a stable reduction of non-HDL cholesterol. (1) Data limitation in the derivation cohort. (2) Strong clinical assumption that treatment effects are sustained over a much longer term than has been studied in clinical trials. Reported using R but codes not available
Silva, R. (2016), Observational-Interventional Priors for Dose-Response Learning. In Advances in Neural Information Processing Systems 29. Single intervention, Treatment dose (continuous), Conditional effect Infant Health and Development Program (IHDP) Continuous A(1) and additionally: It is possible to collect interventional data such that treatments are controlled (1) Computation complexity. (2) Have not discussed at all the important issue of sample selection bias. (3) Generalisability issue. Code available from OLS
Van Amsterdam, W. A. C. et al. (2019). Eliminating biasing signals in lung cancer images for prognosis predictions with deep learning. npj Digital Medicine, 2(1), 1-6. Single intervention, Discrete choice (0/1), Average effect Lung caner Survival A(1), A(2) and additionally:
An image is hypothesized to contain important information for the clinical prediction task. The collider can be measured from the image.
(1) Provide an example of how deep learning and structural causal models can be combined. Methods combining machine learning with causal inference need to be further developed. Code available from OLS
Alaa, A. M., & Van Der Schaar, M. (2017). Bayesian Inference of Individualized Treatment Effects using Multi-task Gaussian Processes. In Advances in Neural Information Processing Systems 30. Single intervention, Discrete choice (0/1), Conditional effect IHDP & Heart transplantation for cardiovascular patients Continuous/Survival times A(2) (1) No experiments regarding outcome prediction accuracy. (2) The computational burden is dominated by the O(n3) (matrix inversion on line 13 in Alg.1. Code available from authors' website.
Arjas, E. (2014) Time to Consider Time, and Time to Predict? Statistics in Biosciences. Springer New York LLC, 6(2), pp. 189-203 Single intervention, Discrete choice, Conditional effect Acute middle ear infections Survival A(2) and local independence In studies involving real data the computational challenge can become formidable and even exceed what is feasible in practice. NA
Pajouheshnia R. et al. (2020) Accounting for time-dependent treatment use when developing a prognostic model from observational data: A review of methods. Stat Neerl. 74(1). Multiple intervention; Discrete choice (0/1); Average effect Chronic obstructive pulmonary disease (COPD) Survival A(1) and A(2) A very strong indication for treatment will result in structural non-positivity leading to biased estimates of treatment-naïve risk. NA
Sperrin, M. et al. (2018) Using marginal structural models to adjust for treatment drop-in when developing clinical prediction models, Statistics in Medicine. John Wiley & Sons, Ltd, 37(28), pp. 4142-4154. Multiple intervention; Discrete choice (0/1); Conditional effect CVD Binary A(1) and A(2) (1) Have not modelled statistical interaction between treatment and prognostic factors; (2) Did not explicitly model statin discontinuation; (3) Only consider single treatment. Code available from OLS
Lim, B. et al. (2018). Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks. In Conference on Neural Information Processing Systems 32. Multiple intervention; No restriction on treatment choices; Average effect Cancer growth and treatment responses No restriction A(2) NA Code available from OLS
Bica, I. et al. (2020). Estimating Counterfactual Treatment Outcomes over Time through Adversarially Balanced Representations. ICLR 2020 Multiple intervention; Discrete treatment choices; Average effect Treatment response in a tumour growth model No restriction A(2) Additional theoretical understanding is needed for performing model selection in the causal inference setting with time-dependent treatments and confounders. Code available from authors' website.
Xu, Y. et al. (2016) A Bayesian Nonparametric Approach for Estimating Individualized Treatment-Response Curves. Edited by F.Doshi-Velez et al. PMLR , pp. 282-300. Multiple intervention; Discrete treatment choices; Conditional treatment effect (1) kidney function deterioration in ICU; (2) the effects of diuretics on fluid balance. Continuous A(2) NA NA
Soleimani, H. et al. (2017). Treatment-response models for counterfactual reasoning with continuous-time, continuous-valued interventions. In Uncertainty in Artificial Intelligence. Proceedings of the 33rd Conference, UAI 2017. Continuous-time intervention; Continuous-valued treatments; Conditional treatment effect Modelling physiologic signals with EHRs for treatment effects on renal function No restriction A(2), A(3) While this approach relies on regularisation to decompose the observed data into shared and signal-specific components, new methods are needed for constraining the model in order to guarantee posterior consistency of the sub-components of this model. NA
Schulam, P., & Saria, S. (2017). Reliable Decision Support using Counterfactual Models. In Advances in Neural Information Processing Systems 30. Continuous-time intervention; Continuous-valued treatments; Conditional treatment effect Applicable to data from EHR but not restrict to such medical settings Continuous-time; no restriction on data type A(2), A(3), A(4) (1) the validity of the CGP is conditioned upon a set of assumptions that are, in general, not testable. The reliability of approaches therefore critically depends on the plausibility of those assumptions in light of domain knowledge. Code available from authors' website.
  1. Abbreviations for ‘Stated Assumptions’: A(1) Relevant directed acyclic graphs (DAGs) available, A(2) Identifiability conditions (consistency, exchangeability, and positivity; or sequential version of consistency, exchangeability, and positivity for time-varying treatments), A(3) Continuous-time exchangeability and A(4) Non-informative measurement times. Other abbreviations: EHR electronic health record, RCT randomised controlled trial