Digital therapeutics

“Digital therapeutic” names a regulatory status rather than a technology: software as a medical device (SaMD) that claims a treatment effect for a named indication and is therefore obliged to demonstrate it. Inside that frame there is a real mechanism — but it does not live in the code.

Where the active ingredient sits

What acts is the behavioural intervention the software delivers. The firmest example is cognitive behavioural therapy for insomnia (CBT-I), recommended as first-line treatment for chronic insomnia by the American College of Physicians. Its active components are known and physiological: sleep restriction raises homeostatic sleep pressure and consolidates sleep; stimulus control restores the learned association between bed and sleep that hours of wakeful lying in bed have broken; a fixed rise time stabilises circadian phase. All of these are instructions the patient executes between sessions, which makes the therapist mainly a source of scheduling, prompting and sleep-diary review. That is precisely the role software can take over without loss — hence outcomes comparable to face-to-face delivery, and hence the point of the category: not a new active principle, but the removal of a supply constraint on clinicians.

Where the claimed mechanism is not a behavioural protocol but “attention training” or plasticity in general, the evidence is thinner. Transfer from the trained task to the clinical symptom has to be demonstrated separately, and it is usually small.

The control problem

A software trial is hard to blind. A placebo tablet is indistinguishable from the active one; a placebo app is not, because the patient can see they are playing a different game, and expectancy alone moves the self-reported scales that psychiatric endpoints are made of. That forces objective or semi-objective endpoints and active comparators rather than waitlists. The regulatory routes reflect the same seriousness: in the US these clear as devices, usually through De Novo or 510(k); in the EU, MDR 2017/745 places software with a diagnostic or therapeutic function in at least class IIa with a clinical evaluation behind it.

Decay as pharmacokinetics

The category’s signature failure is engagement collapse. Dose here equals modules completed, so attrition is functionally equivalent to stopping a drug, and it is the main reason efficacy in a trial and effectiveness in practice diverge. Retention for openly available digital interventions is frequently in the low single-digit percent by a few weeks. That, not the interface, is the real engineering problem of the field: holding the dose.

Category versus mechanism

The honest line runs here. If the product delivers a protocol with a known mechanism and has been tested against an active comparator in the indication it claims, it is a therapeutic. If it logs numbers, reminds and motivates without a defined protocol and without a trial, it is a wellness service, and the word “therapeutic” in its name describes a market rather than an action.