Definition, product form, and regulatory boundaries
Digital therapeutics clinical trial services design, conduct, and report human studies of products that deliver structured therapeutic interventions through software. Within WSIC 6.5.1, sleep applications may include cognitive behavioral intervention for insomnia, circadian adjustment, adherence training, or other programs whose software content and interaction produce a therapeutic effect. A product may be used alone or with a drug, device, or clinical service. Deliverables include evidence strategy, protocol and statistical design, software configuration, remote recruitment and consent, data flows, monitoring, safety management, analysis, reporting, trial registration, and disclosure support.
“Digital therapeutic” is a product and industry term; it does not automatically determine legal status in every jurisdiction. Software intended to diagnose, treat, or prevent disease may be medical device software, while general wellness education, sleep logging, telehealth platforms, and digital tools used only to collect trial data do not become digital therapeutics for that reason. The service differs from the preceding general sleep-device trial entry, the next entry on human consumption studies for functional foods, ordinary software testing, and routine care. A trial also cannot replace software verification, cybersecurity, human factors, or a quality system, and it cannot guarantee authorization, reimbursement, or benefit.
Intended use, mechanism, and clinical evaluation
Development begins by fixing the intended use: target disease and stage, population, standalone or adjunctive role, prescription or nonprescription mode, user, platform, treatment duration, outputs, contraindications, and proposed claims. Clinical evaluation should connect three bodies of evidence: a valid clinical association between the software output or intervention and the target clinical state; analytical or technical validation that inputs, algorithms, and outputs perform to specification; and clinical validation that meaningful outcomes occur in the target population under real use conditions. Literature and neighboring products may provide context but do not replace direct evidence for the studied product version.
Therapeutic action may arise from content, sequencing, personalization rules, feedback, reminders, goal setting, rewards, therapeutic alliance, or human support. The protocol should identify core components and optional services. Insomnia studies should distinguish short-term symptom improvement, sleep-diary change, daytime function, and long-term relapse; circadian interventions also require light exposure, schedules, shift work, and time zones to be documented. When benefit depends on clinician coaching, hardware sensors, or concomitant treatment, neither evidence nor labeling should attribute the combination result to software alone.
Software version, algorithms, and trial design
The investigational product is jointly defined by its application, server, content library, decision rules, model, supported operating systems, interfaces, and language version. Release date, build, configuration, and feature flags should be recorded. A pivotal trial should lock a version representative of marketing. Security patches, content revisions, and algorithm updates require assessment of their effects on risk, blinding, enrollment, exposure, and endpoints, with prospective rules for a pause, renewed consent, bridging, or stratified analysis. Training, tuning, validation, and clinical test datasets should remain separate; optimization based on unblinded outcomes cannot be followed by treating all participants as exposed to one version.
Comparators may include usual care, a waitlist, health education, an attention-matched digital control, an active treatment, or another suitable option, but each controls expectations, contact time, and co-interventions differently. A digital sham may have effects of its own or be recognizable, so credibility, risk, and ethics require evaluation. Randomization should limit contamination within households, therapists, or communities. Because participants and implementers may be difficult to blind fully, blinded endpoint assessment, objective rules, and prespecified analyses become important. Exploratory pilots and confirmatory trials must not blur their objectives, version maturity, or strength of inference.
Endpoints, engagement, and estimands
Endpoints should match the treatment claim and experiences important to patients. A sleep digital therapeutic may assess insomnia severity, diary-based latency and nocturnal wakefulness, total sleep time, sleep quality, daytime function, quality of life, and durability. Polysomnography, actigraphy, or wearable data should be used only when fit for the question and adequately validated. App opens, page views, clicks, and course completion measure process and engagement; they are not clinical benefit. App-store ratings, marketing retention, and feedback from self-selected users cannot replace controlled clinical outcomes.
Primary endpoints, clinical meaning, assessment windows, multiplicity, and subgroups should be defined prospectively. The estimand should address uninstalling, discontinuation, cross-arm use, escalation to human therapy, new medication, device replacement, account sharing, and life events. A digital treatment “dose” may combine lessons, exercises, time, and sequence, but completion may affect benefit and also be affected by early benefit, burden, and health, making post hoc high-versus-low use comparisons biased. Missing data should distinguish unanswered assessments, offline devices, synchronization failures, and a true absence of events, with sensitivity analyses testing assumptions.
Remote conduct, human factors, and equitable access
Remote recruitment, electronic consent, identity checks, online assessment, home use, and remote follow-up should have responsibilities, locations, time zones, contacts, and emergency referrals mapped in the protocol. The team must confirm understanding rather than record a click alone, verify that remote measurement is fit for the population, and ensure that sites, local clinicians, support staff, and technology vendors know when to escalate a medical issue. Automated reminders and frequent contact may improve completion while changing intervention intensity. Trial-supplied devices, connectivity, coaching, and rapid support should be identified if they will not belong to the marketed model.
Human-factors evaluation should cover critical tasks such as registration, login, password recovery, navigation, entry of sleep information, interpretation of advice, notification settings, pausing, help seeking, and withdrawal, while considering nighttime fatigue, cognitive load, sensory or dexterity limitations, and caregivers. Bring-your-own-device designs introduce variation in screens, sensors, operating systems, and networks, while standardized devices can reduce real-world representativeness. People lacking a smartphone, stable connectivity, digital literacy, or a majority language should not be silently excluded. Recruitment funnels, technical barriers, support needs, completion, and outcomes across groups should be reported.
Safety, privacy, cybersecurity, and clinical escalation
Safety planning should cover symptom worsening, daytime sleepiness and driving risk, mania or psychiatric symptoms, suicide or self-harm risk, incorrect advice, delayed care, overreliance, unauthorized access, and service interruption. Lack of physical contact does not mean lack of harm. Adverse events should be linked to version, content, time of use, algorithm output, notifications, human interaction, and concomitant treatment. Risk stratification, crisis messaging, and escalation should fit the population. Unvalidated keyword or score triggers cannot replace clinical judgment, and emergency protection cannot depend on the user remaining online.
Data governance should distinguish research-essential data, product operations, and optional analytics and apply data minimization, clear consent, role-based access, encryption, retention, deletion, cross-border controls, and vendor oversight. Provenance and transformation should be traceable across sensors, apps, third-party components, interfaces, cloud services, and analytic repositories. Threat modeling, vulnerability handling, patches, backup, outage recovery, and incident response span the lifecycle. A privacy policy or de-identification claim does not eliminate re-identification and inference risks or authorize undisclosed advertising or model training with trial data.
Statistical reporting, lifecycle change, and claim limits
Analysis and reporting should present all important outcomes by randomized assignment and prespecified estimand, while disclosing the recruitment funnel, activation, use trajectories, version exposure, technical failures, human support, contamination, loss to follow-up, and safety events. A large volume of digital data does not create independent samples or stronger evidence; repeated clicks and many nights of recording cannot inflate the participant count. Negative, uncertain, and subgroup findings should be reported, with registration, results disclosure, and retention completed as applicable. Algorithms, code, databases, and analytic environments should allow key results to be reproduced.
Content, platforms, models, interfaces, and network environments continue to change after launch. A predetermined change plan should bound proposed modifications, validation methods, impact assessment, and transparent communication, with changes outside that boundary reevaluated. Conclusions apply only to the studied indication, population, version, language, platform, support model, treatment course, and follow-up. Short-term self-reported improvement does not establish durable function, relapse prevention, or health-care utilization benefit. Trial services do not provide individual treatment advice, portray engagement or exploratory findings as efficacy, or guarantee a risk-free product, universal response, or regulatory and payer acceptance.