Safety-first Architecture for Personalised Digital Therapeutics

Photo by Jair Lázaro on Unsplash

The challenge of safe algorithmic personalisation

Digital therapeutics provide evidence-based treatments delivered directly through software interventions. As these technologies mature, health systems and software developers are looking for methods to adapt treatment content to individual patient responses in real time. However, tailoring interventions automatically using artificial intelligence introduces distinct clinical risks. Current machine learning approaches for digital therapeutics often lack dedicated safety frameworks. Furthermore, clinical teams face a scarcity of long-term outcome data and must contend with unpredictable model behaviours. When an algorithm automatically modifies a therapeutic intervention, clinicians and health informaticians require assurance that the software will not introduce harm or deviate from established clinical standards.

How the SAFE_DTx framework operates

To address these challenges, researchers have proposed SAFE_DTx, an architectural framework designed to place patient safety at the core of digital therapeutic software. The framework integrates established principles of predictive modelling with constrained decision-making. Its defining structural feature is the deliberate separation of response forecasting from clinical decision logic, creating a clear division of responsibilities within the software system.

The proposed architecture consists of two distinct modules:

• AI feedback prediction module: This component analyzes patient data to forecast short-term responses to prospective software interventions.
• Constrained planning module: This component selects the next therapeutic action while applying explicit safety constraints to restrict intervention choices.

By decoupling prediction from planning, the system enforces clear safety guardrails. This design allows digital therapeutics to dynamically adapt to individual patient feedback in real time while remaining strictly within evidence-based safety limits. In addition, separating these components enhances transparency, making the system's internal decision-making process easier for technical and clinical reviewers to audit.

Preliminary findings from computer simulation

To evaluate whether this structural concept is viable, the researchers conducted an in silico evaluation. This study design tests software behaviour within a simulated computer environment rather than in a clinical trial with human participants. Consequently, the findings provide evidence of preliminary architectural feasibility rather than real-world clinical effectiveness or safety.

Within the simulated environment, the SAFE_DTx framework demonstrated promising technical results. Compared to standard baseline intervention strategies, the framework achieved higher levels of simulated patient engagement. Importantly, the system recorded zero safety violations during the testing runs. The constrained planning module successfully prevented the predictive models from breaching the predefined safety parameters established by the research team.

What the framework means for health informatics

For health informaticians, software engineers, and digital health educators, this conceptual framework provides a practical model for managing algorithmic uncertainty. The authors note that the architecture "aligns with emerging regulatory emphasis on AI transparency and patient safety." By building hard limits directly into the software design, developers can create adaptive interventions without relying on predictive models to infer clinical safety rules on their own.

For health systems and low-resource settings, structural approaches like SAFE_DTx offer a prospective pathway for testing adaptive software tools safely. Enforcing explicit safety boundaries helps ensure predictable software behavior, which is particularly critical in settings where continuous human oversight of digital interventions may not be available. Educators can also use this modular model to teach students how to separate predictive machine learning outputs from safety-critical decision logic.

Cautions and open questions

While the architectural proposal is well-reasoned, the nature of the study design means its findings carry clear limitations:

• Simulated testing: The evaluation was conducted entirely in silico. Computer simulations cannot fully replicate real human behaviour, clinical complexity, or unpredictable individual patient responses.
• Lack of long-term validation: Because long-term outcome data for adaptive digital therapeutics remains scarce, short-term response predictions in simulation may not accurately reflect long-term clinical safety.
• Unproven clinical utility: Achieving preliminary technical feasibility in a simulation does not guarantee that the framework will prevent adverse events or maintain user engagement in prospective human clinical trials.

Until the SAFE_DTx framework is evaluated in real-world clinical studies with actual patients, it remains a promising architectural concept rather than a clinically proven safety solution.

Reference

Rim D. SAFE_DTx: Safety-First Framework for AI-Driven Personalization in Digital Therapeutics. JMIR medical informatics. 2026. doi:10.2196/78202. PMID: 42789929.

This research digest summarises a recently published article. It was drafted with AI assistance from the article's abstract and reviewed by the author before publication.