Precision medicine is reshaping the landscape of medtech by tailoring treatments and diagnostics to each patient’s unique clinical profile. As innovative devices and diagnostics enter the market, it becomes critical to understand not only whether they work, but how much, and for whom. Causal inference offers a robust statistical framework for distinguishing true treatment effects from associations, a challenge that is particularly pronounced in observational studies and real-world data settings.
Regulatory bodies are increasingly recognising the value of causal inference methods for generating real-world evidence. In the EU, MDR Annex XIV Part B and MDCG guidance 2020-7/2020-8 outline post-market clinical follow-up (PMCF) requirements that often rely on observational data; MDCG 2020-7 names real-world evidence analysis as a PMCF method and requires manufacturers to justify the suitability of the methodology chosen. In the US, the FDA’s December 2025 guidance “Use of Real-World Evidence to Support Regulatory Decision-Making for Medical Devices” acknowledges the role of rigorous observational study designs in supplementing pivotal trial findings and in supporting expanded indications for use [1]. Both frameworks are in motion: in December 2025 the European Commission published a proposal to revise the MDR and IVDR, reformulating Article 61 and several post-market surveillance provisions, which is currently progressing through the ordinary legislative procedure.
This blog post explores how causal inference methods can be applied in the medtech context to enhance evidence gathering in precision medicine.
Establishing Causality in Medtech Research
Randomised controlled trials (RCTs) have long been considered the gold standard for determining causality in terms of treatment efficacy. However, RCTs are not always feasible or ethical, especially in a context where apps or devices may be iteratively improved or used in real-world settings. Observational data from electronic health records (EHRs), wearable sensors, and remote monitoring systems are increasingly important sources of evidence. In such settings, confounding variables can obscure the true effect of an intervention.
Causal inference methods, such as propensity score matching (PSM), have been widely adopted to address these challenges. Austin [2] provides an extensive review of propensity score methods, demonstrating how matching patients on observed covariates balances treated and untreated groups. It must be stated plainly that PSM addresses only measured confounders. In EHR and registry data, many of the most important confounders are frequently unrecorded. These might include: frailty, disease severity, and clinician judgement about who is a viable candidate for treatment. No amount of propensity adjustment touches unmeasured confounding at all. Because residual confounding cannot be eliminated, it should be quantified: quantitative bias analysis, such as calculating E-values [3], and negative control outcomes [4] allow researchers to assess how much unmeasured confounding would be required to overturn a finding. Directed Acyclic Graphs (DAGs) provide a complementary framework for stating assumptions about these confounding structures explicitly, before an estimation method is chosen.
Instrumental variable (IV) analysis is especially useful when unobserved confounders may bias results. In medtech, natural experiments (such as variations in the adoption rates of a new diagnostic tool across different hospitals) can serve as instruments. Angrist and Pischke [5] discuss how IV methods infer causality in applied econometrics. Provider- and facility-level practice variation is an established instrument family [6]. For an IV to be valid, four core assumptions must hold: independence (the instrument is as-good-as-randomly assigned with respect to unmeasured determinants of the outcome), relevance (the instrument strongly predicts treatment), exclusion (the instrument affects outcomes only through treatment), and monotonicity (if the target is a local average treatment effect). Facility-level instruments are vulnerable on two of these, and the failures are distinct. Adoption tracks hospital size, teaching status and case mix, which are unmeasured determinants of outcome correlated with the instrument, an independence failure. Adopting hospitals may also deliver different co-interventions alongside the device, opening a second pathway to the outcome—an exclusion failure. Both require an explicit argument alongside balance diagnostics [7].
Integrating Causal Inference with Machine Learning
Machine learning (ML) has further enriched the causal inference toolkit. Traditional statistical models can be combined with ML techniques to handle high-dimensional data for evaluating heterogeneous treatment effects (HTE) across different patient subgroups. Causal forests, which extend the causal tree framework of Athey and Imbens [8] by averaging across many trees, estimate conditional average treatment effects (CATEs) without requiring the moderating covariates to be nominated in advance [9]. (These covariates must still be measured and included in the model.) This suits the question of which patient subgroups derive most benefit from a device intervention, which conventional interaction models can only answer for moderators specified a priori. Because subgroup effect estimation is prone to overfitting and false discovery, causal forests use honest splitting, employing disjoint subsamples for tree-splitting and effect estimation, to support valid inference. Like PSM, these ML methods require unconfoundedness and positivity to yield valid causal estimates. When rigorously applied, HTE analysis can help ensure that treatments are optimally matched to those most likely to benefit.
Structural equation modelling (SEM) is also used to map out the causal pathways between device interventions and clinical outcomes. By delineating both direct and indirect effects, SEM offers a structured account of how innovations in medtech may influence patient care. Decomposing direct and indirect effects requires strong assumptions [10]: no unmeasured exposure-outcome confounding, no unmeasured mediator-outcome confounding, no unmeasured exposure-mediator confounding, correct functional form, and no exposure-induced confounding of the mediator-outcome relationship. The last of these is easily violated in device pathways—for instance, where a device alert triggers an intermediate clinical action that both changes management and independently affects the outcome. SEM is therefore best presented as a structured hypothesis about mechanism, with its assumptions stated, rather than as a demonstrated causal cascade.
Enhancing Post-Market Surveillance of Medical Devices
Causal inference plays a crucial role in post-market surveillance. Observational studies conducted in post-market settings can suffer from biases that obscure a device’s true performance over time. Techniques such as difference-in-differences (DiD) analysis and targeted maximum likelihood estimation (TMLE) [11] are employed to assess the ongoing impact of the technology. DiD is an identification strategy that differences out time-invariant unobserved confounding under an assumption of parallel trends. TMLE is a doubly-robust estimation approach that combines machine learning with causal inference to reduce bias, though, like PSM, it addresses measured confounders only.
For example, DiD is a standard method for evaluating health policy changes [12]. It can be used to evaluate the impact of remote patient monitoring technologies on hospital readmission rates by comparing readmission trends before and after the implementation of monitoring programmes across different health systems. Such analyses can help attribute observed changes to the medtech intervention rather than to broader secular trends, provided the parallel trends assumption is argued and evidenced with pre-period data rather than simply asserted.
From Diagnostic Accuracy to Personalised Treatment Strategies
A compelling application of causal inference in precision medicine is its ability to inform personalised treatment strategies. A practical example can be found in the evaluation of AI-driven diagnostic tools. In a study evaluating the CheXNeXt deep learning algorithm for chest radiograph diagnosis, researchers compared the algorithm’s performance against practicing radiologists across 14 pathologies using a dataset from a single institution [13]. The reference standard was the majority vote of three cardiothoracic radiologists. CheXNeXt outperformed radiologists on atelectasis, while radiologists were significantly better on cardiomegaly, emphysema, and hiatal hernia, with no significant differences on the other 10 pathologies.
This study reports pathology-level variation in accuracy, not subgroup-level variation. Heterogeneous treatment effects concern variation in the causal effect of an intervention on an outcome across patient subgroups. The bridge from measurement to HTE can be built by specifying the intervention and the outcome: if the algorithm’s output changes clinical management, and management changes patient outcomes, then the causal effect of deploying the algorithm may vary by patient subgroup. This occurs partly through accuracy and partly through how clinicians respond to it. Subgroup-level variation is the relevant analogue, though it was not reported in this study. Defining this intervention and outcome transforms a measurement question into a defensible HTE question: one that shows where along the chain from detection to management to patient outcome the benefit is being realised or lost. That is a different and more actionable finding than knowing where the algorithm’s accuracy is weakest.
Challenges and Future Directions of Causal Inference in MedTech
While the promise of causal inference in precision medicine is immense, challenges remain. The real constraints in these analyses are unmeasured confounding, positivity violations, and measurement error in EHR-derived covariates. A dominant current framework for addressing these constraints is target trial emulation [14]. This approach forces researchers to explicitly specify the eligibility criteria, treatment strategies, and outcomes of a hypothetical trial, thereby organising the observational analysis around a valid causal estimand. This framework specifically addresses endemic device registry issues such as immortal time bias [15] and prevalent user bias.
Methodological development continues. Double/debiased machine learning, for instance, allows flexible estimation of nuisance parameters while retaining valid inference for the treatment effect [16]. But the returns to better estimators are limited when identification is the binding constraint. The larger gains in medtech are likely to come from better study design: prospectively planned registries capturing the covariates that actually drive treatment selection, pre-specified target trial protocols, and routine reporting of sensitivity analyses.
Causal inference stands as a critical pillar in advancing precision medicine in the medtech arena. By enabling researchers to untangle complex causal relationships from observational data, these methods allow the benefits of innovative devices and diagnostics to be quantified and attributed, with their assumptions stated and their limitations known. By embracing causal inference alongside rigorous target trial emulation, medtech companies can accelerate the development of truly personalised solutions that improve clinical outcomes and redefine patient care.
References
[1] U.S. Food and Drug Administration. (2025). Use of Real-World Evidence to Support Regulatory Decision-Making for Medical Devices. Guidance for Industry and Food and Drug Administration Staff, December 2025. Docket FDA-2023-D-4395.
[2] Austin, P.C. (2011). An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies. Multivariate Behavioral Research, 46(3), 399–424.
[3] VanderWeele, T.J., & Ding, P. (2017). Sensitivity Analysis in Observational Research: Introducing the E-Value. Annals of Internal Medicine, 167(4), 268-274.
[4] Lipsitch, M., Tchetgen Tchetgen, E., & Cohen, T. (2010). Negative Controls: A Tool for Detecting Confounding and Bias in Observational Studies. Epidemiology, 21(3), 383-388.
[5] Angrist, J.D., & Pischke, J.S. (2009). Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press.
[6] Brookhart, M.A., Wang, P., Solomon, D.H., & Schneeweiss, S. (2006). Evaluating Short-term Drug Effects Using a Physician-Specific Prescribing Preference as an Instrumental Variable. Epidemiology, 17(3), 268-275.
[7] Brookhart, M.A., & Schneeweiss, S. (2007). Preference-based Instrumental Variable Methods for the Estimation of Treatment Effects. International Journal of Biostatistics, 3(1).
[8] Athey, S., & Imbens, G. (2016). Recursive Partitioning for Heterogeneous Causal Effects. Proceedings of the National Academy of Sciences, 113(27), 7353–7360.
[9] Wager, S., & Athey, S. (2018). Estimation and Inference of Heterogeneous Treatment Effects Using Random Forests. Journal of the American Statistical Association, 113(523), 1228–1242.
[10] VanderWeele, T.J. (2015). Explanation in Causal Inference: Methods for Mediation and Interaction. Oxford University Press.
[11] van der Laan, M.J., & Rubin, D. (2006). Targeted Maximum Likelihood Learning. The International Journal of Biostatistics, 2(1), Article 11.
[12] Dimick, J.B., & Ryan, A.M. (2014). Methods for Evaluating Changes in Health Care Policy: The Difference-in-Differences Approach. JAMA, 312(22), 2401–2402.
[13] Rajpurkar, P., Irvin, J., Ball, R.L., et al. (2018). Deep Learning for Chest Radiograph Diagnosis: A Retrospective Comparison of the CheXNeXt Algorithm to Practicing Radiologists. PLoS Medicine, 15(11), e1002686.
[14] Hernán, M.A., & Robins, J.M. (2016). Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. American Journal of Epidemiology, 183(8), 758-764.
[15] Suissa, S. (2008). Immortal Time Bias in Pharmacoepidemiology. American Journal of Epidemiology, 167(4), 492-499.
[16] Chernozhukov, V., Chetverikov, D., Demirer, M., et al. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1-C68.
