Positioning
TweenMe starts from a straightforward observation: digital twin and AI tools are today inaccessible to most of the professionals who need them most. Not for lack of data — many hold a substantial data asset and precise research questions — but because these tools presuppose an algorithmic mastery that the physician, clinical researcher or domain expert neither has nor should need to acquire.
TweenMe is addressed to the domain expert, an AI layperson, who has research questions and a data asset capable of answering them.
It is a universal digital twin generator in Do It Yourself mode — designed so that the domain expert remains in control of their research question, their data and their results, without any algorithmic intermediary. The promise is twofold: accessibility (no AI prerequisite required) and speed (weeks from raw data to a reliable digital twin, versus months with traditional approaches).
TweenMe does not produce a model that gets installed and forgotten. It produces a digital twin with managed quality and a lifecycle controlled end-to-end — versioning, update traceability, drift detection, inference auditability. The domain expert remains the owner of their system at every stage.
The universality of the architecture has been demonstrated across heterogeneous domains — predictive medicine, clinical oncology research, urban transport. Wherever a professional holds structured data and a precise question, TweenMe can operate.
Transparency note: TweenMe has been implemented in 4 paying POCs in real operational environments (TRL 6–7). The clinical and territorial interfaces described (GHT dashboard, predictive alerts) are under development as part of PREDICARE — they are not part of the core engine currently deployed. MDR certification is planned but not yet engaged.
Technical architecture
TweenMe is the technological response to a precise question: how to build, from heterogeneous industrial data of variable quality, a specific and dynamic computational representation sufficiently faithful to support relevant predictive decisions?
This question is generic — it applies to any domain where a professional holds an instrumented data asset. In the specific case of health and life sciences data, it takes a particularly demanding form.
The HDLSS problem — High Dimensionality, Low Sample Size
Healthcare data presents a structural characteristic that defeats most classical machine learning approaches: it is high-dimensionality, low sample size (HDLSS). A patient can generate thousands of variables (biological, clinical, omics, behavioural) — but the cohort available to train a predictive model is often limited to a few dozen or hundreds of subjects.
In this context, standard models overfit or fail to extract generalisable patterns. The data is structurally rich but volumetrically thin — which inverts the usual logic of classical deep learning approaches.
TweenMe module
Smart Data Fertilizer
In response to the HDLSS problem, TweenMe integrates a data enrichment module — the Smart Data Fertilizer — which augments the volume of available data in a controlled manner, preserving the internal statistical structures of the real cohort. The objective: to enable an AI to learn the internal patterns of a dataset too narrow to be directly exploited.
This module is distinct from naive synthetic data generation — it operates a structurally informed fertilisation, validated on clinical oncology trials (ISPOR Glasgow 2025, top 5% of posters).
How TweenMe works
Starting from an analysis of your data asset and your research question, TweenMe operates in four sequential, fully guided steps — with no algorithmic expertise required on your part.
Representation
Knowledge graph construction
TweenMe builds a knowledge graph and a purpose-fit ontology for your data asset, calibrated to maximise the presentation of useful information with respect to your research question. Your data is no longer viewed generically — it is structured with intention and perspective specific to your objective.
Enrichment — Smart Data Fertilizer
Data augmentation
TweenMe enriches your asset along two complementary axes:
- Adding dimensions — multimodal fertilisation by injecting "donor" datasets selected for their structural coherence with your asset
- Increasing sample size — generation of structurally informed synthetic data to resolve the HDLSS problem and enable the AI to learn the internal patterns of your cohort
Modelling — Experimental design
Model selection & optimisation
Based on your research question, TweenMe selects from 25+ deep neural network architectures the most appropriate ones, then runs an automated experimental design:
- Sizing and depth parameterisation of selected networks
- Hyperparameter optimisation to maximise accuracy under cost and energy constraints
- Automatic comparison of candidate configurations to identify the best performance/resource trade-off
Deployment
Encapsulation & production
TweenMe encapsulates the finalised model in a dedicated webapp, allowing you to use it for inference without any additional technical infrastructure. You remain the owner of your model, your data and your results — with complete end-to-end lifecycle traceability.
Design principles
TweenMe is designed around three non-negotiable principles, drawn directly from the RAISE Framework and the deployment constraints of regulated environments:
- Native interoperability — FHIR R4 / HL7 as exchange standard, not as an adaptation layer. Data enters and exits in standardised formats without proprietary transformation.
- Data sovereignty — certified HDS hosting, French localisation, no health data transiting outside the contractually defined sovereign perimeter.
- Auditability by design — every twin update, every predictive model inference, every generated alert is traced and reconstructible for clinical and regulatory audit purposes.
Collection & Ingestion
Modelling
Hosting
Data governance
Interfaces ¹
Compliance
¹ Clinical and territorial interfaces (GHT dashboard, predictive alerts, regulatory export) are under development within the PREDICARE project. They are not part of the core engine currently deployed in POCs.
Architecture universality
TweenMe is not a healthcare-specific tool. Its individual digital twin generation architecture rests on generic principles — multimodal trajectory modelling, real-time updating, predictive inference — applicable to any sufficiently instrumented complex system.
Two non-healthcare use cases have already demonstrated this universality:
Smart Transportation
TweenMe has been applied to urban mobility and transport flow problems, as a substitute for traditional simulation models (agent-based models, O-D matrices). Results outperformed traditionally deployed models on predictive accuracy metrics, with significantly reduced implementation time. This use case demonstrates that the individual digital twin architecture can model complex sociotechnical systems beyond the medical domain.
Synthetic population generation — Oncology
A synthetic population generation algorithm developed within the TweenMe framework, applied to a therapeutic trial in lung cancer, was presented at ISPOR Glasgow 2025 (International Society for Pharmacoeconomics and Outcomes Research) and ranked top 5% of posters at the conference.
This result validates the approach on a high-stakes regulatory use case — synthetic population generation for clinical trials is subject to particularly stringent statistical and regulatory validation requirements (ICH E6, 21 CFR Part 11). The top 5% ranking by an international peer jury constitutes a significant independent validation.
Scientific recognition
Top 5% — ISPOR Glasgow 2025
Synthetic population generation algorithm applied to a therapeutic trial in lung cancer. Evaluated by international peer jury — International Society for Pharmacoeconomics and Outcomes Research, Glasgow 2025.
Technology Readiness Level (TRL 1–9)
Current position: TRL 6–7 — 4 paying POCs completed · Validated in real operational environments
The PREDICARE pilot phase (GHT cohort) constitutes the step towards TRL 8–9 — full system validation in an operational environment, a prerequisite for MDR certification and industrial-scale deployment.
RAISE Framework application
TweenMe is the technological building block of PREDICARE — as such, it is subject to the same RAISE requirements, with specific constraints relating to its status as a Software as Medical Device (SaMD).
Regulatory Architecture
MDR 2017/745-oriented design from initial architecture — risk classification, technical documentation, clinical evaluation, ISO 13485 quality management system. EU AI Act: high-risk AI system classification (Annex III).
Interoperability Standards
Native FHIR R4 — no proprietary adapter. Interoperability with leading hospital information systems (EHR, LAB, RIS) without institution-specific development.
Safety & Operational Validation
Clinical validation protocol on the PREDICARE cohort. All predictive model outputs are decision-support tools — the final clinical decision remains under medical responsibility. Graceful degradation mechanism in case of unavailability.
Explainability & Ethics
Two-level explainability: clinician level (clinical reasons for the alert) and technical level (variable contributions to the predictive score). Granular patient consent. Ethics committee integrated into the PREDICARE governance.
Relationship with PREDICARE
TweenMe is to PREDICARE what an engine is to a vehicle — a central technological building block, necessary but insufficient on its own. PREDICARE provides the territorial infrastructure, clinical governance, funding model and regulatory framework within which TweenMe can function as a deployable system.
Conversely, TweenMe is the condition of possibility for PREDICARE — without a digital twin generator, the territorial programme has no computational layer to build on.
This interdependence is structural, not contingent. It is documented in the PREDICARE Memoir v3.