Target trial framework: A topic on everyone's mind at the 42nd ISPE Annual Meeting

Adigens Health helps pharmaceutical and rare disease companies design real-world evidence strategies built around the target trial framework: from early development planning through HTA submission. Get in touch at info@adigenshealth.com.

Adigens Health was at the 42nd ISPE Annual Meeting (formerly ICPE) in Milan recently, co-authoring sessions on the target trial framework (TTF), ECAs and measurement error.

By our count, the TTF was mentioned in approximately 65 session and poster titles – spread across a range of therapeutic areas, including oncology, diabetes, chronic kidney disease, pregnancy and vaccines.

The pre-conference training days ran several dedicated courses including ‘Target Trial Emulation (TTE) for Causal Inference in Practice’, ‘Defining and Estimating the Effects of Dynamic Treatment Strategies using Real-World Data’, and ‘Learning from benchmarking against reference trials with real-world data’.

The resources you need for flawless Target Trial Emulation 

Adigens co-authored the symposium: From Inception to Reporting: The Resources You Need for Flawless Target Trial Emulation, with Xabier García de Albéniz, Miguel Hernán, James McAuley, and Aidan Cashin. The authors presented recent developments within the TTF, discussed how to articulate your research question by specifying a target trial, walked the audience through the TARGET reporting guidelines and presented common mistakes in TTE and how the tools presented can help. 

Externally Controlled Trials (ECT) for Evidence Generation in Regulatory Decision-Making: challenges, emerging frameworks and regulatory use cases 

The externally controlled trials symposium, led by Andrei Barbulescu from EMA with Miguel Hernán among the co-authors, brought together regulatory experts from PMDA, MHRA and the FDA to discuss external control arms (ECAs) in practice. Hernán covered the use of the TTF to construct an ECA that complements a single-arm trial, how to diagnose emulation failures through benchmarking and negative control outcomes, and the view that if you can run a randomized control arm, you always should. The panel also discussed the issues of contemporaneity, time-zero and eligibility misalignment, and the lack of any agreed threshold for how much borrowing a given regulatory decision can tolerate when assessing ECAs. 

Identify, Quantify and Act: advancing Methods for Reducing Bias from Measurement Error 

The symposium on measurement error, led by Rosa Gini, with García de Albéniz among the co-authors, worked through the importance of identifying where exposure or outcome misclassification can be introduced and how to control for it. This session made the point: a perfectly specified target trial can still be undone by exposure or outcome misclassification, because the target trial framework only controls design bias. 

Hot Topic Session: Acetaminophen (paracetamol) and Autism: Evidence interpretation and communication 

The hot topic session covered maternal acetaminophen (paracetamol) use in pregnancy and the risk of autism spectrum disorders. The panel summarized the literature, discussed key methodological considerations, including sibling designs, and paid particular attention to communication and translating complex, uncertain and conflicting data to inform regulatory decisions and clinical recommendations in pregnancy.

That translation problem is particularly challenging in pregnancy research as randomized trial data on treatment effects in pregnancy is rarely available, so post-approval observational studies end up carrying most of the evidentiary weight. That burden makes the case for the TTF clearly: specifying causal questions in pregnancy, emulating a target trial against the data sources that can actually support it, and recognizing the design assumptions that don't transfer cleanly from non-pregnancy TTE work.

Other topics of interest 

The agenda in Milan was packed enough that no single recap could cover it all, but a few other topics stood out. Clone-censor-weighting, the technique used to emulate treatment strategies that are not distinguishable at time zero without introducing immortal time, turned up across nine separate sessions. There was a mini session dedicated to self-controlled designs, and negative control outcomes featured across a range of sessions, including a dedicated symposium on recent developments and best practice. 

AI and large language models applied to real-world data had a track of its own, covering EHR data cleaning, clinical note analysis, synthetic data, and ML risk prediction. GLP-1 receptor agonist safety got its own mini-session plus a long tail of posters. COVID-19 vaccine safety surveillance had a full lightning-session block and in a session titled "One Size Does Not Fit All," presenters pushed back on average treatment effects altogether, arguing for causal-forest and personalized-effect methods over a single population-level estimate.

Our take 

The Target trial framework ran through many sessions, posters and pre-conference training at the 42nd ISPE conference in Milan. Across oncology, diabetes, chronic kidney disease, pregnancy and vaccines, it was not presented as a new method. Most sessions assumed it rather than introduced it, and the methodological discussion was largely about how to apply it rather than whether to. Its extensions, such as the TARGET reporting guidelines and the clone-censor-weighting are following the same path.

This changes what is worth paying attention to, because using the TTF is no longer a distinguishing feature of a study, but specifying it well is. The specification determines what question is being asked, and errors introduced at that stage, like eligibility criteria that cannot be mapped to the available data, treatment strategies that are not sufficiently well defined, or a time zero that does not align with eligibility and assignment, are not recoverable at the analysis stage.

An important take away is that even if the perfect TTE specification removes design bias, it does not address unmeasured confounding, measurement error or missing data. For unmeasured confounding specifically, it can only be detected, using negative control outcomes for example, and quantitative bias analyses. 

Specifying the target trial well and addressing the limits of the available data are separate obligations, and both are required.