Digital therapy for panic disorder worked far better with a clinician attached

The commercial case for digital therapeutics rests on a single premise: software scales where clinicians do not. A meta-analysis published this month in the Journal of Anxiety Disorders tests that premise directly for panic disorder, by separating trials in which a clinician was involved from those in which the app worked alone [s1].

What was pooled

The analysis included 40 randomised controlled trials of digital interventions for panic disorder published up to March 2025 [s1]. Eligible trials enrolled adults with a primary diagnosis of panic disorder, with or without agoraphobia, and compared a digital therapeutic against either an active control — therapist-led treatment or treatment as usual — or a passive one, meaning a waitlist or no treatment [s1].

Three outcomes were measured: the Panic Disorder Severity Scale (PDSS), the Agoraphobic Cognitions Questionnaire (ACQ) and the Body Sensations Questionnaire (BSQ) [s1]. The first captures how severe the panic disorder is; the latter two capture the fear-related thoughts and bodily-sensation fears that cognitive behavioural models treat as the maintaining mechanism.

Random-effects meta-analyses were run alongside subgroup analyses, sensitivity analyses and mixed-effects meta-regressions, with comparator type, guidance format, intervention modality and region as moderators [s1].

The split

On panic severity, self-guided digital therapeutics produced a moderate effect (Hedges' g = 0.31, 95% CI 0.05 to 0.68) [s1]. Clinician-guided versions produced a considerably larger one (g = 0.95, 95% CI 0.44 to 1.46) [s1].

On the cognitive outcomes, the gap becomes categorical rather than a matter of degree. Only clinician-guided interventions produced statistically significant improvements: ACQ g = 0.46 (95% CI 0.15 to 0.76) and BSQ g = 0.67 (95% CI 0.30 to 1.05) [s1]. Self-guided formats showed negligible effects on the same measures — ACQ g = 0.11 and BSQ g = 0.27 [s1].

So an app on its own reduced how severe panic disorder was, and did close to nothing to the catastrophic beliefs about bodily sensations that the therapy is nominally designed to change.

The authors' reading is domain-specific rather than dismissive: well-structured self-guided interventions can address symptom domains involving panic frequency and physiological distress, while clinician involvement exerts a notably stronger influence on cognition-related outcomes [s1]. Their conclusion for design and policy is that digital therapeutic products should be matched to the mechanistic pathway through which change is expected to occur [s1].

Why the confidence intervals matter here

The self-guided PDSS estimate has a lower bound of 0.05 — only just clear of zero. The clinician-guided estimate ranges from 0.44 to 1.46, which is wide. These are pooled estimates across 40 heterogeneous trials with different apps, durations and comparators, and the guidance-format comparison is a subgroup contrast rather than a head-to-head randomisation. No trial in this pool randomised the same app with and without a clinician; the difference is between studies, which is a weaker design for causal claims than the effect sizes suggest.

The same pattern outside mental health

A scoping review published at the end of the month, covering digital therapeutics for prediabetes, documents how routinely a human is embedded in these products [s2]. Of 21 included studies — 17 of them randomised controlled trials, searched to 10 March 2025 — the delivery modalities were smartphone apps in 14 (67%), human coaching in 13 (62%), messaging tools in 9 (43%), wearable devices in 9 (43%) and web platforms in 3 (14%) [s2].

Human coaching appears in nearly two-thirds of these prediabetes digital therapeutics [s2]. Whatever is being evaluated in that literature, it is frequently not software alone.

The review also found that behavioural grounding is often implicit: the most common theories were social cognitive theory, the theory of planned behaviour and the transtheoretical model, but 11 of the 21 studies applied behaviour change techniques without explicitly stating a theoretical framework [s2]. The most-used techniques were self-monitoring of behaviour (19 of 21), instruction on performing the behaviour (16 of 21), goal setting (15 of 21), information about health consequences (15 of 21) and unspecified social support (11 of 21) [s2].

And the outcomes measured skew toward the proximal: metabolic and body composition measures in 19 of 21 studies, glycaemic control in 17, cardiovascular risk and physiological function in 16, behavioural and cognitive indicators in 11 — and comprehensive health outcome measures in just 2 [s2].

What this means for how these products are described

If clinician guidance is doing much of the work, then the economics of digital therapeutics look different from the pitch. The panic disorder analysis is explicit that its findings bear on regulatory framing and scalable deployment [s1] — the question of whether a product evaluated with a clinician in the loop should be authorised, reimbursed or marketed as a standalone.

Neither paper says self-guided tools are useless. The panic analysis found a real, if modest, effect on symptom severity without a clinician [s1]. The claim it undermines is the stronger one: that removing the human leaves the therapeutic effect intact.

This article does not offer treatment advice, and neither study addresses which option is appropriate for any individual.

What to watch

Trials that randomise guidance directly — the same app, with and without clinician contact, in one population. Until those exist, the size of the guidance effect rests on between-study comparison, which is the weakest link in an otherwise well-specified analysis [s1].

Sources

• Clinician guidance in digital therapeutics for panic disorder: Meta-analytic dissection and implications for regulatory framing and scalable deployment — Journal of Anxiety Disorders, 2025-09-19
• Application of Behavioral Science in Digital Therapeutics for Individuals With Prediabetes: Scoping Review — Journal of Medical Internet Research, 2025-09-29