Interventions delivered by digital technologies – such as the Internet, smartphone apps, and chatbots – represent a promising means of increasing the availability of psychological treatments. A mature yet evolving body of evidence indicates that digital interventions can effectively target many psychiatric symptoms, whether delivered as stand-alone self-help programs, adjuncts to traditional in-person treatment, or relapse prevention tools1. Yet, digital interventions are characterized by high rates of early dropout or disengagement. In some trials, nearly half of participants never initiate use of their assigned digital intervention, or discontinue use entirely within the first 10 days2. Problems with engagement are even more profound in real-world settings. Early dropout is a critical concern, because it limits the potency of a digital treatment due to insufficient therapeutic content exposure, and undermines the validity of clinical trial findings. Preventing dropout is therefore a key priority. Clinicians can play a central role in preventing early disengagement when digital health tools are integrated into a patient's care plan. A critical point for clinician intervention is at the beginning of treatment, as this is when early engagement trajectories can be established. Premature dropout from digital interventions typically occurs when patients have unclear expectations regarding optimal use, and limited understanding of how these tools can support treatment goals3. Moreover, without clear plans in place at treatment initiation, early difficulties may be interpreted as evidence that the digital intervention is burdensome, ineffective, or poorly suited to the user's needs, further contributing to disengagement. Clinicians can mitigate this risk through intentional intake dialogue that clarifies when and how digital tools should be used, aligns demands with patient capacity, and defines the platform's role within the broader care plan. This may include explicitly articulating how digital interventions serve complementary functions, such as providing psychoeducational material outside of sessions or structured homework tasks that reinforce therapeutic work. Clinicians can also proactively anticipate likely barriers to sustained use, and work collaboratively with the patient to engage in early problem-solving aimed at addressing these challenges as they emerge, rather than waiting until patterns of non-use are established. The earlier phase of treatment also represents a critical window for clinician support, as risk of dropout is highest during the first few weeks2. Disengagement during the initial treatment phase can be due to difficulty integrating the digital intervention into daily routines, uncertainty about whether initial effort is translating into progress, or skepticism about how the homework activities connect to in-session therapeutic work3. Clinicians can address these risks by front-loading support through brief, structured check-ins during early weeks of use. These check-ins can focus on collaboratively troubleshooting how the digital health tool fits into the patient's daily life, helping patients recognize early signs of skill acquisition even in the absence of clinical change, and explicitly linking completed digital homework tasks to in-session goals. By reinforcing relevance and strengthening the perceived value of continued use, early check-ins can support sustained engagement during this period of heightened dropout vulnerability. The success of these engagement strategies, however, depends fundamentally on clinicians possessing adequate digital literacy. Patient-reported usability issues – including limited technological competency, navigation difficulties, or complications interpreting clinically relevant digital data – are often-cited contributors to early dropout4, and therefore require clinician capacity to anticipate and address them. Without sufficient familiarity with the digital health tools they recommend, clinicians may struggle to troubleshoot technical barriers, help patients extract meaningful feedback from their data, or ensure that these interventions are being used optimally. Building digital literacy through credible training curricula5 can equip clinicians with the skills needed to address patient-reported usability challenges, thereby ensuring that digital health tools are used in ways that support, rather than hinder, treatment goals. While these clinician-led strategies offer immediately actionable ways to reduce early dropout, they rely extensively on clinician time, judgement, and manual monitoring, highlighting the need for complementary approaches that can support more efficient identification of disengagement risk. Achieving this efficiency calls for robust data-driven methods that are capable of accurately identifying which patients are likely or unlikely to disengage prematurely. Such predictive capabilities could support more precise allocation of retention efforts and inform clinical decision-making on the suitability of digital interventions for individual patients. Developing reliable, data-driven predictive algorithms capable of supporting automated personalization at the individual-patient level necessitates access to large volumes of data drawn from multiple sources. Algorithms trained on rich, heterogenous data are better positioned to detect subtle patterns and interactions that predict disengagement than those relying on sparse or single-modality inputs, which may overlook critical context-specific dropout risk factors and exhibit reduced predictive stability6. Digital health platforms are uniquely positioned to support this approach, as they can capture continuous multimodal data streams, including textual data derived from completed homework activities, usage metadata, passively sensed behavioral data, and temporally dense, high-frequency symptom ratings collected via repeated monitoring surveys. These data streams may capture distinct and complementary prognostic signals related to dropout that are not readily observable using traditional clinical baseline assessments (e.g., disruptions in sleep regularity inferred from passive smartphone data may signify accumulating fatigue that undermines engagement). Crucially, collection and use of passively sensed and textual data requires explicit, informed user consent, which must be obtained transparently and in a manner that respects patient privacy and autonomy. Consent procedures should be designed to minimize burden and avoid undermining engagement, as overly complex or intrusive data requests could contribute to dropout1. Thus, multimodal data capture approaches should be implemented selectively, and carefully evaluated for their acceptability and impact on user engagement. Despite the demonstrated value of integrating multimodal data streams to predict patient outcomes in other areas of psychiatry7, this approach has not yet been systematically applied to the problem of predicting dropout from digital interventions. Existing work predicting early dropout has largely examined one or two data streams in isolation, with the strongest evidence supporting early usage metadata and textual features, which demonstrate modest predictive accuracy8. While informative, these studies are typically proof-of-concept, conducted on single datasets, and rarely subjected to external validation, thereby limiting their clinical relevance and readiness for real-world deployment. Progress in this field demands systematic evaluation of whether multimodal data modelling can meaningfully improve the prediction of early dropout in digital mental health. Advances in predictive performance should be accompanied by rapid progression beyond proof-of-concept toward prospective and externally validated models tested in large, diverse samples, as this will be needed to establish generalizability and suitability for real-world deployment. Achieving this at pace will require coordinated, global collaboration among digital mental health researchers, including shared data infrastructures, harmonized data collection protocols, consensus on feature definitions, and improved standards for model reporting. Once multimodal predictive models are robustly validated, their clinical value can be assessed within implementation-focused trials that test whether algorithm-guided, just-in-time support can more effectively prevent early disengagement than non-adaptive digital interventions6. In conclusion, early dropout remains a critical barrier to realizing the potential of digital mental health interventions. Clinicians are central to encouraging patient engagement through supportive strategies implemented early in treatment. Harnessing data-driven predictive algorithms that integrate multimodal information can offer a means to augment these strategies with a level of precision far exceeding what manual monitoring can feasibly achieve. If validated and implemented effectively, such algorithms could be embedded directly within digital health platforms to provide real-time risk alerts that prompt timely clinician intervention. By combining clinician expertise with data-driven prediction, this integrated approach holds promise for preventing early dropout and maximizing the therapeutic potential of digital interventions.
Jake Linardon (2026) studied this question.