In Summary
- Longitudinal synthetic data moves beyond single-timepoint patient snapshots to simulate realistic healthcare trajectories over time, including visits, diagnoses, treatments, and outcomes.
- Data sources encompass EHRs, clinical registries, administrative claims, clinical trial follow-up and extensions, and wearable or sensor streams.
- Key challenges are managing irregular event timing, variable record lengths, complex temporal dependencies, and rare clinical events across multi-year histories.
- Longitudinal synthetic data can support research into progressive and chronic diseases, enable privacy-preserving collaboration, and inform clinical trial design, including the development and assessment of synthetic comparators.
- Aindo sees longitudinal generation as a natural extension of its work in synthetic data and evidence generation, bringing together clinical relevance, privacy, and a certified data-processing framework.
A single photograph is not a biography
Picture two assignments for a portrait painter. The first: paint one convincing portrait of a person you have never met, from a description of what they tend to look like. The second: paint fifty portraits of that same person, one for every year of their life, so that laid out in order they read as a believable biography: the child becoming the teenager becoming the adult, each face still recognisably itself, ageing in ways a viewer would accept as a real life rather than fifty unrelated faces sharing a name.
For cross-sectional synthetic patient data, a model learns the statistical “look” of a population and generates a new, artificial patient as a single snapshot. With longitudinal synthetic data, instead of one snapshot, the model generates a patient’s course through the health system: visits, diagnoses, treatments, and events, so that it holds together both moment to moment and as a story. A response attributed to a treatment must follow that treatment, not precede it; a patient’s five-year follow-up must fit consistently with their baseline and intervening history. That second assignment is considerably harder than the first, and it opens an important opportunity for clinical research.
What is “longitudinal” data?
With longitudinal data, we mean any type of data recorded for one individual on multiple occasions or time points. They can be recorded in different ways, come from different sources, and serve different purposes. In healthcare, longitudinal data come mainly from the following sources:
- Electronic health records (EHRs): the densest and messiest kind, encompassing the visits, diagnoses, lab results, prescriptions, and procedures a patient accumulates in the ordinary course of care. Records are created whenever care is delivered, which can mean fifty visits in ten years for one patient and two for another, at intervals set by illness, habit, and access rather than a study protocol.
- Clinical registries: a more curated cousin, with structured, disease-specific records, such as cancer, rare-disease, or cardiovascular registries. Typically built by clinicians or consortia to track how a condition and its treatment evolve, they may offer more consistent variables than a general EHR, while follow-up completeness varies.
- Administrative and claims data: records from the operational and financial side of healthcare, including insurance billing, reimbursement codes, and hospital discharge data. They can offer a broad picture of when someone interacted with the health system, even when the clinical detail of what happened is thin.
- Clinical trial follow-up and open-label extension data: patients are followed according to a protocol, with specified measurements collected at planned times. These data are closer to the portrait painter’s regularly spaced portraits than most real-world sources, providing a clearer temporal framework, although missed visits, treatment changes, and dropout still need to be addressed.
- Patient-reported outcomes and wearable or sensor data: symptom diaries, quality-of-life questionnaires, continuous glucose monitors, and heart-rate or activity trackers. These sources range from periodic questionnaires to high-frequency streams, with gaps whenever a patient stops reporting or a device comes off.
Unlike clinical trial protocols that specify planned visit schedules, most observational healthcare data follow an irregular, highly individualised cadence. This temporal irregularity and sequence dependence present a substantial challenge to generative modelling, requiring the synthetic process to learn both the clinical events and their timing structure.
Why a timeline is so much harder to reproduce than a snapshot
A single synthetic patient record only has to be plausible once. A synthetic patient history has to be plausible at every point along the way, and it has to be plausible as a sequence. Three things make this genuinely difficult, not just bigger:
- Order and timing carry as much information as the facts themselves. A model generating a single row has to get what is true about a patient right. A model generating a trajectory also has to get when it becomes true and place it in a sequence consistent with the available clinical evidence. Both the sequence and the timing of clinical events are meaningful: a model has to capture dependencies across time, not just what a typical event looks like in isolation.
- No two patients’ timelines look alike. A table has one row shape. A set of medical histories does not: some patients are seen twice, others fifty times, over wildly different spans, at intervals driven as much by access to care as by biology. Generating that kind of irregular, variable-length data remains an active area of research.
- The rarest, most important events are the hardest to place correctly. A late-onset side effect or a relapse years after remission is already difficult to represent because it is rare. The model also has to place it correctly in a person’s timeline, not just acknowledge that it happened.
Put simply: reproducing one photo convincingly is a modelling problem. Reproducing a life story convincingly is a modelling problem and a logic problem and a timing problem, stacked on top of each other.
What is the state of the art?
The central challenge has always been to give models enough context to generate what comes next. Earlier sequential and generative approaches already captured aspects of patient history; more recent methods have expanded the ways in which longer and more complex dependencies can be represented. A sequence can look convincing visit by visit and still become inconsistent when viewed as a whole, which is why remembering the relevant past matters.
One important advance has been improving how models represent what has happened so far before deciding what comes next. This is also a core idea behind the large language models that now write text: rather than generating each word independently, they use the preceding context to decide what follows. Applied to a patient’s record instead of a sentence, the same principle means generating the next diagnosis, lab result, or treatment in light of the patient’s earlier history, rather than treating each visit as an isolated event. Recent studies illustrate this direction through the generation of clinical-event sequences and the forecasting of longitudinal measurements.
Different models build this “memory” in different ways, from recurrent architectures to transformers and other generative approaches. The common objective is to treat a patient’s record not as a set of disconnected facts, but as an ordered story in which what happens next should make sense in light of what came before.
Despite these technical advances, the field still lacks a widely adopted, standardised framework for comparing longitudinal synthetic data across applications; quantitative evaluations already exist, but the measures and benchmarks used vary substantially. To establish trust and clinical utility, synthetic trajectories should be assessed across key dimensions: clinical realism, temporal fidelity, and trajectory coherence, combining quantitative tests with clinical expertise and validation against real outcomes.
A new shift: general-purpose models alongside bespoke ones
Many applications in this field have relied on purpose-built models, trained for a particular dataset or disease area; an emerging direction is to train broader models on larger collections of patient histories. The idea echoes the logic behind general-purpose language models: learn patterns from a broad base, then investigate how that foundation can serve more specific questions.
A model that has absorbed patterns across a large patient population may offer a useful starting point for a disease-specific application, although the benefit must be demonstrated rather than assumed. Work such as Delphi-2M shows how generative models can produce possible future disease sequences at population scale. The opportunity is not simply a larger model, but a broader starting point that can be assessed against the needs of an individual research programme.
It is worth being precise about what this does and does not solve. A bigger, more broadly trained model does not automatically fix timing, logic, or rare-event problems; whether it offers a stronger foundation for a particular use must be tested. Privacy, clinical coherence, and suitability for a specific purpose should remain explicit requirements, whichever model is chosen.
Impact on the life sciences
Longitudinal synthetic data holds transformative potential across several key areas of the life sciences:
- Research into progressive and chronic diseases: Conditions such as Alzheimer’s, Parkinson’s, and multiple sclerosis require understanding change over time. Generative modelling has already been used to simulate Alzheimer’s progression from longitudinal observations. For research teams, the opportunity is to explore multi-year scenarios using existing evidence, examine differences between patient groups, and prepare hypotheses before launching a new study, without mistaking simulated follow-up for new clinical observations.
- Disease-progression modelling: Beyond standard cohort analysis, artificial generation of longitudinal histories offers a complementary perspective. Established frameworks often represent pathways through predefined health states; generative models can instead learn patterns from sequences of observations and generate possible continuations from an existing patient history. This brings the focus closer to individual trajectories: not just whether a patient progresses, but how measurements and events may change along the way. Integrated alongside robust health-economic, causal, and epidemiological frameworks, dynamic simulations provide a useful lens for exploring long-term clinical questions.
- Privacy-preserving cross-border and inter-institutional collaboration: For multi-year patient histories, the aim should be to preserve useful clinical patterns while assessing and managing disclosure risk, with privacy safeguards evaluated as part of the generation and release process. Empirical privacy tests and formal guarantees are distinct approaches, and should be described accurately for the application concerned. Combined with appropriate governance, this creates an opportunity to make longitudinal evidence more accessible to collaborative research.
- Clinical trial planning and synthetic comparators: By modelling longitudinal profiles, sponsors can explore inclusion and exclusion criteria, candidate endpoints, and follow-up schedules before committing to a protocol. Where synthetic cohorts are intended to augment a control group, their suitability must be established for the particular study rather than assumed from their realism alone. The practical ambition is clear: better-informed design choices, more efficient use of research resources, and a more purposeful use of the data patients have already contributed.
Aindo’s position
Aindo brings more than generative AI technology to longitudinal data generation: it combines a dedicated team with experience working alongside leading Italian clinical and research centres. Its clinical network includes primary institutions grounding its work in the realities of specialist care and clinical research.
This experience is backed by an established workflow that takes clinical data from ingestion and harmonisation through synthetic generation, validation for analytical utility, bias, and privacy, and delivery of analysis-ready cohorts. For pharmaceutical companies, the value is not another modelling tool to integrate, but a partner that connects the clinical data, technical expertise, and validation needed to address a specific research question.
Aindo applies this approach to the challenges of longitudinal generation: selecting representative source data, fitting the modelling strategy to the disease and available history, and evaluating whether the resulting trajectories preserve the timing, progression patterns, and outcomes relevant to the study. The focus is on what makes a simulated patient journey useful for a development programme: understanding disease heterogeneity, exploring endpoint behaviour, informing follow-up strategies, and assessing the suitability of synthetic comparators.
Supported by Europrivacy, ISO 9001, and ISO 27001 certifications, Aindo combines its clinical-data expertise with an established framework for data protection, quality, and information security. This combination defines its proposition to pharma: dedicated expertise, experience with clinical institutions, and a structured path from complex healthcare data to research-ready evidence. Longitudinal generation brings these capabilities together around a more demanding question: not simply what a patient looks like today, but how their disease evolves over time.
Let’s shape the future of healthcare together
Are you interested in exploring how longitudinal synthetic data can support your clinical research, trial design, or collaborative studies?
Let’s connect and explore your vision.