Digital mental health services promise to tailor care to each individual, but what counts as personalisation is often unclear. Marketing teams (and researchers) use the term to describe superficial features (for example, letting users choose a background colour or selecting a programme based on age and gender). Meanwhile, researchers developing adaptive interventions are working with sophisticated data streams and algorithms to match treatments to an individual’s evolving needs.
This post untangles these concepts for an informed but non‑technical audience, drawing on academic literature and some of my own work. This is an issue I have often seen, and a recent review of a research grant has promoted me to write this article.
What counts as personalisation?
In health‑technology research, personalisation refers to adjusting an intervention so that it becomes meaningfully relevant to a specific person rather than a broad segment. A scoping review of digital mental health interventions defined personalisation as a process that changes a system’s functionality, interface or information content to increase its personal relevance. This can involve tailoring the timing, content and intensity of support based on an individual users data, adapting to feedback, or using models that learn over time.
Researchers distinguish several dimensions of tailoring. Static tailoring uses information gathered at a single point (for example, a baseline questionnaire, or even a first name) and does not adjust/adapt again. Dynamic tailoring collects data over time and adapts accordingly; interventions may change the type, timing or intensity of support as a user's context or mental state changes. Surface tailoring responds to observable behaviours - think frequency of app use - while deep tailoring responds to determinants of behaviour such as motivation, self‑efficacy or social norms. Personalisation can be user‑determined, where the individual user chooses the features they want, or computer‑determined, where algorithms decide what to deliver.
In practice, true personalisation requires four key elements:
- Adaptive interventions: systems that adjust the type, timing or dosage of support based on real‑time data. Just‑in‑time adaptive interventions (JITAIs) deliver tailored content when the user is receptive; they rely on sensors and machine‑learning models to infer context and determine the moment of delivery. For example, a JITAI might recognise patterns of stress in sensor data (wearable or phone) and prompt a grounding exercise.
- Dynamic feedback loops: personalisation is an iterative process. The intervention monitors how a user responds and adjusts accordingly. This could involve active feedback (self‑reported mood, symptom check‑ins, mood diary) and passive data (movement, sleep, phone usage). When we say passive data this can come from a mobile phone or a wearable device. Digital phenotyping, uses passive data from smartphones or wearables, aims to reduce burden on participants while providing continuous behavioural measures.
- Context awareness: personalisation considers the user’s environment and circumstances, not just their demographic characteristics. It may draw on location data (GPS), time of day, social context or recent behaviours to provide relevant support. My recent work on remote measurement and digital phenotyping illustrates this; combining passive smartphone and wearable data with active questionnaires can detect changes in symptoms and enable targeted support.
- Behaviour change techniques: effective digital interventions embed evidence‑based behaviour change techniques (goal setting, self‑monitoring, feedback on performance) and adapt these techniques to the user’s progress. Our DrinksRation app uses weekly mood and mental‑health questionnaires to tailor push notifications and, previously, SMS messages that encourage diary use, suggest alternative behaviours and guide goal setting.
True personalisation versus superficial tailoring
Many apps labelled as 'personalised' offer only simple segmentation. Common misconceptions include:
- Demographic segmentation: categorising users by age, gender or occupation and delivering the same static content to everyone in that segment. In a commentary for BMJ Military Health we cautioned that interventions should focus on the individual rather than treat groups based on demographic characteristics. Age‑ or gender‑based tailoring may alter imagery or examples but rarely adjusts the therapeutic strategy.
- Static rule‑based tailoring: using a one‑off questionnaire to assign users to a programme. This may improve relevance initially but cannot respond to changes over time. Without ongoing data collection, it is not possible to deliver timely support when a user’s needs shift.
- Content libraries mislabelled as personalised: some apps claim to personalise simply because they offer a large number of modules and let users choose what to read or interact with. Allowing choice is valuable and key to positive change sin behaviour but does not constitute personalised guidance. It places the burden on users to navigate their care.
- Marketing buzzwords: features like 'AI‑powered' chatbots or 'customised' meditations are often touted without evidence. Many commercially available mental health apps incorporate no sensors and rely on pre‑programmed interactions; a review found that only about 1% of marketplace apps use smartphone sensors and only ~2% have published research evidence.
Such superficial tailoring may improve appeal but does not adjust treatment based on individual data or evidence. By contrast, adaptive interventions use detailed feedback and context to modify support continually. There is a significant on efficacy and outcome.
Evidence and examples: lessons from research
A lot of my work provides real‑world examples of personalisation. In the DrinksRation RCT, participants completed weekly mood assessments. The app algorithm used these responses to personalise push notifications and text messages encouraging diary use and suggesting alternative behaviours. This dynamic tailoring improved engagement and, in a help‑seeking military veteran population, reduced alcohol consumption. We then extended this work to use a simple AI approach, a behaviour change model, and personalised push notifications to further drive changes in behaviour and alcohol consumption.
We also did work on remote measurement with smartphones and wearables, and it was great as it demonstrated the feasibility of passively collecting behavioural data. We found ~92% of participants provided passive data via their devices. Combining passive data with active ecological momentary assessments allowed prediction of symptom escalation and opened the door to adaptive interventions delivering support when needed.
In another article, we emphasised that digital technologies can gather fine‑grained data and that algorithms can recognise behavioural patterns. But we cautioned against treating users as generic age or occupational cohorts and argued for individualised care. We also highlighted emerging evidence that personalised messages delivered via SMS and push notifications, informed by behaviour change theory, improve outcomes in areas such as alcohol misuse and physical activity.
More broadly, reviews of digital mental health interventions reveal a gap between research prototypes and market offerings. A large analysis of 293 commercially available apps for anxiety and depression found that just over half referenced an evidence‑based framework in their app-store descriptions, and only 6% had published evidence supporting efficacy.
Most apps neglect advanced smartphone features; only about 1% integrate sensors for digital phenotyping. Evidence quality across digital psychiatry research is generally low to moderate, with small sample sizes, few RCTs and potential publication bias. Thus, while the potential for personalisation is enormous, most commercial apps lack the dynamic, data‑driven adaptivity that underpins true personalised care.
Why most apps don't deliver meaningful personalisation
Several barriers prevent digital mental health apps from achieving genuine personalisation.
Limited data and superficial algorithms: True personalisation requires longitudinal, multimodal data (symptoms, behaviours, context) to train adaptive algorithms. Commercial apps often rely on static self‑report questionnaires because they lack access to passive sensor data or because incorporating sensors increases development costs and regulatory complexity. As a result, they cannot adjust interventions to momentary changes or detect early signs of relapse.
Evidence and evaluation challenges: The evidence base for digital mental health apps is weak. Even research trials (including my own) suffer from small sample sizes and limited randomised controlled designs; the overall quality of evidence is low to moderate. There is also little consensus on how to measure engagement, adherence or attrition.
A 2025 consensus group highlighted that terms like usage, engagement and adherence are defined inconsistently across studies and that raw usage data are rarely reported. Moreover, increased engagement does not necessarily translate into better outcomes; few studies have examined dose-response relationships, and some interventions require only short exposure to be effective.
High attrition and low engagement further limit personalisation. Digital interventions must gather ongoing data, but real‑world retention is poor. A review of unguided digital mental health interventions, the median programme usage rate in research studies was over four times higher than in real‑world settings, indicating that trial participants often receive extra support not available to typical users. Without sustained engagement, algorithms cannot learn or tailor interventions effectively. This last point is very important to consider.
Data quality and burden: Adaptive systems depend on accurate data. Passive sensing reduces user burden by collecting data in the background, but not all participants value or understand the need for continuous monitoring. Digital health data literature warns that if patients do not experience value in generating and sharing data, they will not comply; likewise, clinicians may find unsolicited data burdensome and unhelpful. Privacy and security concerns also influence adoption; misunderstanding how data will be used or shared undermines trust. There is a risk that personalisation efforts become intrusive, leading to 'feedback fatigue' or anxiety.
Ethically, personalisation raises questions about autonomy, fairness and data ownership. The value‑sensitive design framework for youth digital mental health notes that personalisation should enhance empowerment but must balance requests for user feedback with an uninterrupted user experience (a great article exploring this is here). Overly burdensome data collection can discourage use, particularly among young people with low technological or health literacy. Cultural sensitivity is also essential (and often forgotten); interventions must account for diverse beliefs and expectations to avoid reinforcing stigma.
Deployment and scale: Even when adaptive interventions perform well in trials, scaling them into routine care is challenging. Integrating personalised apps into health services requires training, workflow changes and assurance that algorithms are safe, equitable and interpretable. Digital phenotyping literature notes that implementation science must address stakeholder preferences and regulatory requirements to ensure scalability and sustainability. Many health systems lack the infrastructure to handle large volumes of digital data or to integrate them with electronic health records. Economic models for funding personalised digital care are still evolving.
Why true personalisation is hard
Developing genuine personalisation is not just a technical challenge; it is a multifaceted problem involving data, ethics and evaluation.
- Data quality and burden: Adaptive algorithms require high‑quality, multi‑dimensional data. Passive sensing can reduce participant burden, yet active input (self‑reports, mood ratings) remains necessary to interpret behaviours. Balancing data richness with user burden is crucial; too many questionnaires lead to disengagement (we found this in one study), but too little data hampers adaptation.
- Ethics and safety: Collecting continuous data raises privacy, consent and data‑protection issues. If users do not understand how their data will be used, they may refuse to share it. Algorithms may reinforce biases present in training data. For vulnerable populations, misclassification or inappropriate recommendations can cause harm. Ethical frameworks emphasise empowerment and autonomy; personalisation should not manipulate users or exploit their data.
- Evaluation and evidence generation: Demonstrating that personalised interventions improve outcomes is difficult. Many studies lack robust designs; sample sizes are small; and definitions of engagement vary. Without clear metrics, it is hard to decide whether a personalised system is genuinely more effective than a generic one. RCTs of adaptive interventions require complex statistical methods to account for time‑varying treatment effects, further complicating evaluation.
- Deployment and scalability: Real‑world implementation demands integration with clinical workflows, training for practitioners and regulatory approval. Systems must be designed for diverse populations, not only tech‑savvy early adopters. The costs of maintaining adaptive algorithms and updating them to reflect new data can be substantial. Transparent reporting and regulatory oversight are essential to avoid harm and build trust.
Doing personalisation properly: a way forward
To move from hype to reality, researchers, developers and funders should demand more than superficial tailoring (and question what personalisation actually means). Meaningful personalisation involves:
- Co‑design with users and experts: Engage end users in every stage of development to ensure that personalised features meet real needs and respect preferences. Use value‑sensitive design to balance personalisation, empowerment and autonomy.
- Rich, ethically sourced data: Combine passive sensing with targeted self‑report to capture behaviour and context while minimising burden. Use transparent consent processes, explain data usage and provide users with control over what they share.
- Evidence‑based behaviour change: Embed established behaviour‑change techniques and adapt them dynamically. Use theoretical frameworks and measure outcomes beyond engagement, such as symptom improvement and quality of life.
- Adaptive algorithms and JITAIs: Implement machine‑learning models that adjust interventions to momentary states and contexts. Incorporate digital phenotyping to anticipate periods of vulnerability and deliver support at the right time.
- Rigorous evaluation and transparency: Design RCTs with adequate sample sizes and clear definitions of engagement and outcomes. Publish algorithms and evidence, not just marketing claims. Report negative results and adverse events to build a trustworthy evidence base.
- Integration with care pathways: Link personalised apps to health services. Ensure experts can access and interpret digital data (even PDF exports will help), and provide support when users need human assistance. Develop economic models and governance structures to sustain personalised digital care.
Personalisation in digital mental health is not a gimmick; it is a complex, evolving practice that requires real data, sophisticated algorithms, ethical oversight and rigorous evaluation. Most apps fail to deliver meaningful personalisation because they rely on static segmentation, lack access to continuous data and have little evidence to support them.
Let move beyond buzzwords and invest in adaptive, evidence‑based interventions, we can harness digital technology to deliver mental health care that genuinely responds to the needs of individuals rather than the assumptions of marketers.