Explainability as an ethical requirement for digital phenotyping in adolescents

Introduction: The ethical dimensions of digital phenotyping

The ongoing digitalisation of the healthcare sector, especially in the field of mental health care, opens up new possibilities in diagnostics, prevention, and treatment. The development of digital phenotypes, which aim to provide deep insights into psychological conditions and behavioural patterns through the continuous collection of data via smartphones and other wearable devices, seems particularly promising.^1-3 This technological innovation marks a paradigm shift in the treatment of mental illness, as it replaces the traditional episodic observation with a continuous form of monitoring, and thereby possibly enables more precise and timely interventions. In this context, it is especially relevant to look at the development concerning the mental health of adolescents. On the one hand, this is due to the increased burden on this age group, which has been intensified through the COVID-19 pandemic. On the other hand, there is the increasing prevalence of mental illnesses like depression, anxiety disorders, and addictive behavior.^4-6 The technological affinity of many young people seems to offer a good basis for implementing digital health technologies, while at the same time, the special vulnerability and the specific normative status of this age group raise complex ethical questions. The development of digital phenotypes is not only accompanied by technological aspects, but also by ethical, legal, and societal considerations, which need careful philosophical reflection.

The focus of this article is on the question of explainability as a normative requirement for using such technologies with adolescents. More precisely, the question of this paper is what explainability can and should mean in the specific context of digital phenotyping. The developmental psychological characteristics of adolescence, the specific vulnerabilities of this age group, and the constitutive role of digital technologies in adolescent identity formation demand a redefinition of what it means to make algorithmic systems ‘explainable’. This re-conceptualisation – so the central thesis of this paper – does not only transform peripheral aspects of technical implementation, but touches basic questions regarding the relation between technological transparency and adolescence development, between algorithmic categorisation and self-understanding, between epistemic authority and emerging autonomy. While for adult patients technical transparency or outcome-based justification might be sufficient, for adolescents a form of explainability is necessary that takes into account their specific cognitive developmental stages, respects their forming concepts of health and illness, and supports their ability to critically engage with knowledge claims without overwhelming them. Ultimately, this means that the question of explainability of digital phenotypes in adolescence is not primarily a technical or procedural one, but one that leads directly into the heart of what it means to grow up in a digitalised society – and to understand and shape one's own mental health in this process.

Basics of digital phenotypes in mental health care

The conception of digital phenotypes represents a significant innovation in traditional mental health care by causing a fundamental transformation in the way data is collected and interpreted. In contrast, conventional clinical methods, which are based on periodical examinations and subjective self-reports, are considered an outdated procedure. Digital phenotyping allows instead a continuous, contextual and multimodal capturing of behaviour and biomarkers. ⁷ This is why some scholars and clinicians call for an expansion of the use of digital phenotyping.^3,8 This methodological reorientation shows itself in different forms of data collection, which range from active input by the users to passive background data gathering, using various sensors and functionalities of modern technologies. Smartphones and wearables, for example, record movement patterns through GPS tracking, communication behaviour by analysing the frequency and content of messages, and also biometric parameters like voice quality, breathing patterns and reactions of the pupils. ⁹ The collected data is afterwards analysed through algorithmic procedures, whereby more and more artificial intelligence and machine learning are used, in order to recognise complex patterns and to draw clinically relevant conclusions. The potential added value of these technologies for the treatment of mental illnesses in young people comes from several factors: the higher temporal and spatial resolution of the data gathering, the reduction of memory distortions through real-time capturing, the contextualisation of symptoms in the natural life surroundings, and the possibility for early intervention in upcoming crises. Empirical studies prove the validity of digital phenotypes for different mental disorders such as depression, anxiety disorders and addictive behaviour. ¹⁰ In some cases, the precision of the predictions is comparable with or even better than traditional clinical methods.^11,12 The possible fields of application for digital phenotypes in the area of adolescent mental health are manifold and include both diagnostic as well as therapeutic dimensions. In the diagnostic area, they can be used for early detection and the course observation of different psychological illnesses, especially through recognising subtle changes in behaviour or physiological parameters before the appearance of clinically visible symptoms.^13,14 In the therapeutic context, they allow an intervention that is fitted to the individual need, based on real-time data, and so an adjustment of the treatment in a timely manner becomes possible. ¹⁵ The Ecological Momentary Assessment (EMA) method shows the potential of this approach. Through the immediate capturing of thoughts, feelings and behaviours in the natural environment, more precise data is generated. Furthermore, the ability for reflection and the self-awareness of the adolescents are supported.^16-18 At the same time, the integration of digital phenotypes into mental health care leads to a fundamental redesign of the relationship between patients, therapists and technology, whereby traditional understandings of roles and interaction patterns are challenged. The continuous monitoring and algorithmic evaluation of personal data also raise basic questions about privacy, autonomy, and the right balance between technological support and human decision-making. These questions become especially important in the context of the still-developing autonomy of adolescents.^19,20

The normative dimension of adolescence and mental health

Adolescence constitutes a distinct phase of life, whose normative particularity arises from its specific position between childlike dependence and adult autonomy. ¹⁹ This phase of transition is characterised by complex biological, cognitive, emotional, and social developmental processes, which together form the basis for the emergence of a stable identity and the capacity for autonomy. However, the determination of the norm that defines from which age a person is to be classified as adolescent or already as adult presents itself as a challenge. This is due to the fact that the demarcation between childhood and adulthood is fluid and culturally variable. ²¹ Although the biological marker of puberty provides a certain orientation point, interindividual and gender-specific differences must be taken into account, which make a universal definition difficult. Beyond that, adolescence must be understood as a social construct, the shape of which is subject to historical and cultural transformations, and whose demands and developmental tasks are socially defined.^22,23 This conceptual vagueness also becomes visible in the legal and ethical treatment of adolescents, where age-related boundaries are often regarded as pragmatic but insufficient approximations of actual development. In the context of digital phenotyping, the development of the capacity for autonomy is of special relevance, which must be understood as a gradual process and not as a sudden transition occurring at the attainment of a specific age.^24,25

The mental health of adolescents is a multidimensional construct. Its normative importance is derived from two aspects: the intrinsic value of well-being and the instrumental function it has for current and future life perspectives. Mental health must be considered as an integral part of the well-being of the child and as a fundamental precondition for coping with age-specific developmental tasks. It demands special attention and protection. The significant prevalence of mental disorders in this age cohort makes clear the societal relevance of this topic.^5,26 The effects of the pandemic have led to a further worsening of this problem. ⁴ Also concerning are the social gradients in the distribution of mental illness, with adolescents from socioeconomically disadvantaged families and marginalised groups, such as LGBTQ+ youth, being disproportionately affected.^27,28 These health inequalities raise fundamental questions of justice and highlight the necessity for a differentiated understanding of the social determinants of mental health.

The normative conceptualisation of youth and mental health is closely connected with cultural and societal values, which can be reflected in the algorithmic classification systems of digital phenotyping. While Western, individualistic societies tend to prefer autonomy-oriented models of development, collectivistic cultures maybe have different priorities when it comes to defining mental well-being and what is seen as a proper development path.^12,29 This cultural variability raises fundamental questions about how universal digital phenotyping approaches really are and shows that explainability must not only be understood as a technical issue, but also as a problem of cultural translation. In this context, it seems necessary that explanatory frameworks for digital phenotyping reflect culture-specific understandings of normality, pathology, and development, and include mechanisms to adapt accordingly.

The special meaning of explainability in the context of youth mental health comes not only from the general vulnerability of this age group, but from specific developmental-psychological characteristics, which need a qualitatively different form of explainability. While most adults have stable cognitive schemata and experiences to contextualise incomplete or abstract information, youths are in a phase of active construction of their self- and world-understanding.^30,31 This means that algorithmic decisions about their mental health not only influence their current treatment, but can also shape their developing self-perception and their understanding of normality and pathology. The explainability of digital phenotypes is therefore linked to a developmental-psychological perspective – it must not only create transparency, but also support the formation of adequate health concepts and self-images, without determinating them in problematic ways. Exactly because young people are in a phase of increased neuronal plasticity and identity-related exploration, non-transparent algorithmic categorisations of their mental states might have negative effects on their self-conceptualisation, which go beyond the immediate therapeutic effects.

The right to mental health, as anchored in the UN Convention on the Rights of the Child, implies obligations both for parents and for state institutions. A sufficiency-oriented understanding of this right is to be preferred: the goal should not be to achieve the maximum, but a sufficiently good level of mental health. However, the exact determination of this threshold is context-dependent and normatively contested.^32,33 The normative particularity of the adolescent phase and the specific requirements of mental health converge in the question of how to balance appropriately between the respect for autonomy and the responsibility of care. The development of global autonomy, defined as the comprehensive capacity for autonomous life conduct, manifests itself gradually and in domain-specific ways. In certain decision areas, local autonomy may already be developed at an earlier stage. ²⁴ This differentiation is of central importance for the application of digital phenotyping, since complex considerations are necessary here between self-determination, informational self-determination, and health-related well-being. The present investigation deals with the question of how far adolescence can be conceptualised as a social space of protection. According to Joel Anderson and Rutger Claassen, ³⁴ this protective space enables experimentation and risk behaviour, without adolescents having to bear the full consequences. The development of autonomy requires active support by adults as well as appropriate institutional frameworks that guarantee a balance between freedom and care. This complex normative configuration forms the background against which the use of digital phenotypes in adolescents with psychological burdens must be evaluated ‒ especially with regard to the question of explainability as a prerequisite for autonomous decision-making and informed consent.

Explainability as an ethical concept in AI

Explainability is a many-layered normative concept, what in the context of digital phenotyping includes different epistemic, ethical and social dimensions. ²⁰ In the technological discourse, explainability is mostly defined as the ability from a system, to make its function, decision making and results understandable for human users. But this seeming simple definition hides complex philosophical questions about the nature of understanding, the limits of technological transparency and the normative demands on the human-technology-interaction. The explainability of digital phenotyping systems includes different levels: the technical explainability about how algorithms and data processing are working; the epistemic explainability about the validity and trustability of the generated knowledge; the clinical explainability about the medical relevance and what it means for diagnosis and treatment; and the ethical explainability about the normative basic assumptions and values, what are inside the system. ³⁵ These different levels of explainability are in complex relations with each other and can be in tensions, for example when the optimisation of the prediction accuracy goes on the cost of traceability, or when clinical explainability for professionals is in conflict with explainability for youths and parents. The explainability is not understand as binary attribute, what a system has or not, but as gradual concept, that is depending from context and target group and can have different degrees of detailedness, completeness and understandability. ³⁶

Explainability in the context of digital phenotyping with adolescents has normative significance in accordance with several ethical principles. First, explainability functions as a condition for autonomous decision-making, because only a sufficient understanding of how digital phenotyping works, including its implications and potential risks, enables adolescents to make self-determined and authentic choices regarding its use, which is an important ethical principle in medical decision making also for adolescents.^37,38 Second, explainability constitutes the premise for informed consent. In this regard, the developmental requirements for information transfer and comprehension in adolescents must be taken into special account. ³⁹ Explainability is closely linked to the principle of non-instrumentalisation. The realisation of transparent and traceable systems allows adolescents to be treated as autonomous subjects rather than merely as data objects. Fourth, explainability contributes to the building of trust by forming the basis for a trustworthy relationship between adolescents, parents, healthcare providers, and technological systems. ⁴⁰ Fifth, explainability allows for critical reflection and questioning of the values and norms embedded in systems, which makes it possible to uncover and address hegemonic power structures and potential mechanisms of discrimination.

However, it must be accepted that the necessity of explainability for autonomous decision making and informed consent is by far not undisputed. A growing number of scientists are questioning the supposedly self-evident connection between algorithmic transparency and patient autonomy. Adam J. Andreotta, Nin Kirkham and Marco Rizzi, ⁴¹ for example, argue that the demand for full explainability for informed consent puts the bar unnecessarily high, while Sinead Prince and James Edgar Lim ⁴² show that black-box AI does not necessarily undermine patient autonomy, as long as it is properly embedded into the doctor-patient relationship. Jose Luis Guerrero Quiñones ⁴³ goes even further and claims that AI systems can even strengthen patient autonomy by providing more personalised and precise information, without needing full algorithmic transparency. Especially provoking is the position of Suzanne Kawamleh, ⁴⁴ who argues explicit against explainability-requirements for ethical AI in healthcare, showing that human experts also work like ‘black boxes’ without this blocking informed consent. These objections deserve serious attention, because they point to important conceptual distinctions – for example between different forms and levels of explainability, or between the explainability of technical processes and the understandability of clinical implications. But for the specific context of youth mental health, one can argue that these general reflections do not consider enough the special developmental and relational requirements. The question of his paper, though, is not if explainability is needed in general, but how it must be conceptualised and operationalised for this specific population.

The practical implementation of explainable AI systems for digital phenotyping needs a systematic integration of different technical approaches, reaching from local explanation methods – which make single predictions interpretable – up to global interpretation techniques that make the whole behavior of the system more transparent. ⁴⁵ Model-specific methods like Local Interpretable Model-agnostic Explanations (LIME) or SHapley Additive exPlanations (SHAP) can help to make concrete risk assessments for adolescents more understandable by showing which factors – for example changes in movement patterns, communication frequency or sleep behavior – have contributed to a certain evaluation. At the same time, it must be considered that only technical transparency does not automatically lead to better understanding, that's why such explanations should be combined with development-psychology-based communication strategies, which translate complex probabilistic statements into action recommendations that are meaningful and interpretable for adolescents

These ethical dimensions of explainability point to fundamental questions concerning human dignity, the development of autonomy, and the just design of socio-technical systems. Such questions must be taken seriously, especially in the context of vulnerable groups like adolescents with psychological burdens. ⁴⁶ However, the practical realisation of explainability faces significant challenges, both technical and social in nature. The increasing complexity of modern AI systems, particularly those based on machine learning, leads to the so-called ‘black box problem’, where the exact functioning and decisions of the system are no longer entirely understandable—even for the developers themselves. ⁴⁷ This technological opacity stands in fundamental contradiction to the normative demand for explainability and requires innovative approaches to ‘explainable AI’ (XAI), which try to balance transparency and performance. ⁴⁸ The recent critic from Anantharaman Muralidharan, Julian Savulescu and Owen Schaefer ⁴⁹ deserves in this context special attention, because they argue that AI-systems in the healthcare sector should mainly be justified to the patients, and not follow some abstract demand for technical transparency. This position, which makes convincing points, also overlooks the specific normative requirements coupled to explainability in the context of adolescents, who are in a critical phase of building their epistemic autonomy – that means, the ability to evaluate on their own which sources of information are trustful and how to question knowledge-claims in a critical way. The justification of an algorithm decision only by medical professionals, without the possibility to at least partially understand the processes behind, could undermine the development of this critical competence and lead to a problematic epistemic dependency. This doesn't mean that adolescents have to understand the full technical function of machine learning algorithms – even experts hardly can do this. More important is to create age-appropriate forms of explainability, which allow adolescents to get a basic understanding how digital systems come to their estimations about their mental health, which data-types play a role there, and where are the limits and uncertainties of such systems.

Furthermore, the question of the suitable kind and depth of explanation is of crucial importance and must be adapted to the context and the specific audience. For the clinical expert staff, detailed technical explanations can be important, while young people and parents need more everyday-near and action-relevant explanations, which show the main function principles and implications without being too much complicated with technical things. These thoughts show clear that explainability is not only a technical thing, but also a communicative and social challenge, which needs a deep understanding of the needs, abilities and contexts from the involved persons, and makes necessary the development of standards and practices for explainability, which are both technologically possible and ethically correct. ⁵⁰

Explainability as a foundation of relational autonomy and informed consent

The explainability of digital phenotyping systems shows itself in the context of youth mental health as a concept that needs a fundamental redefinition. While traditional approaches of explainable AI are mainly focused on technical transparency and algorithmic traceability, the specific requirements of the youth phase show that explainability here must be understood as a multidimensional, development-sensitive and relational embedded concept. This re-conceptualisation concerns not only the form and the amount of explanations, but also transforms the understanding what it means to make technological systems in the context of vulnerable development phases ‘explainable’. Different than individualistic concepts of autonomy, which see decision-making as mostly inner-psychic ability, the relational approach shows that autonomy is something made in social relations between people.^19,51,52 This theoretical reorientation fundamentally transforms the understanding of explainability: instead of a unidirectional transmission of information, explainability becomes a dialogical process in which meaning is constructed and negotiated jointly – a process, that considers the specific cognitive, emotional and social development dynamics of adolescence and actively includes them in the designing of explanations. The explainability of digital phenotypes must therefore be embedded in the complex relational dynamics between adolescents, parents, practitioners, and technological systems, whereby these relationships both enable and limit the possibilities of explanation and understanding. Developmentally appropriate explainability requires attention to the gradual development of autonomy in adolescents, which varies across domains and must be supported by different forms and levels of explanation. The local autonomy of adolescents in technology-related decisions can be strengthened through tailored explanation strategies that connect to existing knowledge and experiences and make complex interrelations accessible through relevant analogies and examples. At the same time, explainability of digital phenotypes must also support the development of global autonomy by not only enabling punctual decisions, but also contributing to the comprehensive development of competence in dealing with digital health technologies and encouraging critical reflection. Explainability therefore functions as a key resource for autonomy-supporting relationships, in which adolescents are not regarded as passive recipients of technological interventions, but as active co-designers of their own health care. ⁵³ Explainability as a condition for informed consent in the context of digital phenotyping transcends traditional procedural requirements and demands a fundamental reconception of this ethical principle. The epistemic dimension of explainability concerns not only the transmission of factual knowledge about technical functions and medical implications, but also the disclosure of epistemic uncertainties, methodological limitations, and normative assumptions embedded in digital phenotyping systems.

This comprehensive explainability enables adolescents to better understand the scope of their consent and to make considered decisions that align with their own personal values and preferences. But if even the developers of complex AI systems cannot fully comprehend their exact functioning, how can informed consent, in the classical sense, be realised at all? This question points to the necessity of rethinking explainability ‒ not as the complete disclosure of all technical details, but as needs-oriented, context-sensitive, and target group-specific communication of those aspects relevant to decision-making. For adolescents, this means that explanations must be adapted to their developmental level, prior knowledge, and specific informational needs, while both cognitive and emotional dimensions should be taken into account. ⁵⁴ Explainability must further be conceptualised as a continuous process that reflects the dynamic development of both technology and the adolescent capacity for autonomy, and that enables regular review, adaptation, and renewal of consent. This process-oriented conception transforms the understanding of informed consent from a single point event into an ongoing dialogical practice that is embedded in therapeutic relationships and institutional structures.^39,40

The explainability of digital phenotypes is in complex relation with the parental role in the consent process. Traditional concepts of surrogate decision-making must be extended through dialogical and participatory approaches. Relational explainability therefore includes not only direct communication with adolescents, but also the support of parental explanatory competence. This enables parents to convey complex technological and medical concepts in an age-appropriate way and to act as trustworthy interpretive instances. ⁵⁵ At the same time, the requirement for explainability transforms the parental role itself ‒ from authoritarian decision-makers to supporters of autonomy, who help adolescents with processing information, critical reflection, and decision-making. This repositioning requires specific explanatory resources for parents, which consider their own technological and health-related competence and sensitise them to possible tensions between care and respect for autonomy. Triadic explainability between clinicians, parents, and adolescents must reflexively address power asymmetries and allow for flexible participation models, which can vary depending on developmental status and specific context. ⁵⁶ In cases where diverging preferences or values exist between adolescents and parents, differentiated explainability plays a key role. It allows for the disclosure, contextualisation, and potential harmonisation of different perspectives. Explainability thus functions as a bridge between the various actors in the consent process and enables a balance between adolescent autonomy, parental care, and medical expertise that meets the complex normative demands of the adolescent phase and at the same time realises the therapeutic potential of digital phenotyping.

Explainability as a condition for trust and symmetrical transparency

The explainability of digital phenotyping systems constitutes the fundamental requirement for the development and maintenance of trustful relationships in the context of adolescent mental health care. In contrast to simplifying technocratic concepts, which mostly interpret trust as the acceptance of technological authority, an ethically grounded understanding of trust demands a critical reflection on the epistemic, normative, and social dimensions of human-technology interaction. ⁵⁰ Explainability works as a multidimensional bridge between technological systems and human actors, because it creates epistemic transparency, shows basic normative assumptions and opens options for action. In this context, the ability to build trust becomes especially important, because young people are in a phase of strong identity development and increasing critical reflection. This phase is characterised by higher sensitivity for autonomy limitations and manipulation attempts. ⁵⁷ So, explainability must go beyond only technical function and should enable an authentic encounter that respects young people as equal moral subjects and takes their experiences, fears and hopes seriously. The bidirectional character of trust, as Debbie Schachter et al. ⁵³ underline, also shows that explainability which supports trust should not only build trust in the technological systems, but also show trust in the young people themselves – in their ability to understand complex information, reflect critically and make responsible decisions. This reciprocal dimension of explainability transforms the idea of trust in technology from a one-way into a dialogical process, in which trust is built together and must be renewed again and again.

The explainability of digital phenotypes goes beyond classical ideas of transparency, because it not only gives insight into technical processes, but also enables symmetrical transparency relations between all involved persons. In contrast, one-sided transparency demands – which focus mainly on the giving of personal data by the monitored persons – aim at making visible the controlling systems, institutions and decisions. ⁴⁶ The expanded concept of explainability includes different dimensions, which are explained here: First, the algorithmic explainability, about the technical functions and decision logic. Then, the data explainability, about the kind, source and processing of collected data. After that, operational explainability, about how the system is used in clinical workflows and how responsibility is shared between humans and machines. Finally, result explainability, which looks at the validity, reliability and possible bias of the generated knowledge. ³⁵ But this multidimensional explainability faces big challenges, technical and also social. The inner opacity of complex AI systems, especially neural networks, questions the ideal of full explainability and needs new approaches that can balance technical performance and ethical demands for understanding. Also important is that explanations fit to the cognitive ability of the audience. Not only age differences, but also differences in education, tech knowledge and health competence must be considered. All this shows that explainability is not a fixed thing, but a dynamic and context-specific quality that needs continuous adaptation and reflection.

Explainability is a foundation for trust in digital phenotyping systems. It needs new ideas and practices that go beyond traditional transparency and fit to the specific needs of the youth phase. Against the simple idea that more information always means more trust, empirical research shows a more complex picture. Too much technical detail can lead to confusion, insecurity or even mistrust, while too little information can be felt as manipulation or hiding. ⁴⁰ Therefore, explainability must find a careful balance that looks both at cognitive and emotional aspects and offers different explanation levels. Dialogic and participatory explanation strategies are recommended – they allow young people to be active in shaping and evaluating explanations and to express their own information needs. Promoting basic AI literacy in young people, as Davy Tsz Kit Ng et al. ⁵⁸ suggest, is a good foundation for trusting technology. But this should not only be an individual skill, it must be understood as a social practice, embedded in education and health systems. Also, institutional explanation tools like independent evaluations, certifications or ethical guidelines are becoming more important. They act as additional forms of explanation and build a structural frame for trustworthy digital phenotyping. Ongoing reflection and adaptation of these explanation practices together with all actors is a key condition for strong trust relationships – which are crucial for both therapeutic success and autonomy development of young people.

The perspective of adolescents on the sharing of different data types shows complex preference patterns, which are very important for the design of explainable digital phenotyping systems. Empirical studies show that adolescents – even though they in general recognise the benefits of predictive psychiatric services – are very selective when it comes to the types of data they are willing to share. ⁵⁹ While biological and psychosocial data like DNA information or school grades are more often accepted, adolescents are much more hesitant with sharing digital data sources like social media content or Google search history, mainly because of concerns about privacy. In this context, explainability must be seen as a dialogical process that not only makes the technical workings of algorithms transparent, but also reflects the different sensitivity of various data types and gives adolescents possibilities to control how much of their data they want to give.

Explainability as an instrument for power balance and social justice

The explainability of digital phenotyping systems acts as a critical instrument to make visible, analyse and possibly transform inner power structures that shape technological surveillance practices in the context of adolescent mental health. In contrast, technocratic perspectives focus on explainability mainly as a technical challenge. But a critically reflective view requires the analysis of deeper power dynamics that are visible in the design, implementation, and use of digital phenotypes. ²⁰

The methodological explainability addresses fundamental epistemic questions of power: Whose knowledge is included in the system development? It should be found out which phenomena are marked as needing explanation, and which remain invisible. It is to be asked which classification systems and ideas of normality are reproduced by algorithms. This critical dimension of explainability makes it possible to show and problematise asymmetric regimes of visibility, where adolescents are made to objects of technological observation, while the observing institutions, algorithms and their developers mostly stay invisible. ⁶⁰ The conventional phenotyping systems, which are based on a one-directional surveillance logic, are challenged by a multi-directional explainability. This not only allows an understandable algorithmic decision-making for affected persons, but also shows the institutional, economic, and epistemic structures which produce and legitimise these technologies.

The analysis of power-related explainability appears especially important in the context of adolescent autonomy development. The technological intensification of parental and institutional control can strengthen traditional power asymmetries and reduce the free spaces which are essential for identity development. ⁶¹ Reflexive explainability functions in this context as a corrective by giving both adolescents and adults the chance to recognise implicit power dynamics, question them, and find together new balances between care, autonomy, and technological surveillance. The social stratification of digital phenotyping becomes visible in different dimensions of unequal participation and algorithmic discrimination. A critical analysis and addressing of these phenomena needs an extended concept of explainability. The digital gap in technological infrastructure, internet connection and compatible devices creates systematic mechanisms of exclusion, which should be uncovered and addressed by diversity-sensitive explainability. ⁶² Socioeconomic, gender-related, cultural and ethnic inequalities can be reproduced or even made stronger by algorithmic bias, especially if the training data are not representative for the target population or if implicit bias in algorithm development are not properly reflected. ¹²

The diversity-sensitive explainability goes beyond technical precision requirements and touches fundamental questions of representation, participation and epistemic justice: Whose experiences and perspectives are represented in digital phenotyping systems, and which stay invisible? Which culture-specific expressions of psychological distress are recognised algorithmically, and which are not seen? It is to be examined to what extent algorithmic classifications reproduce dominant, Western, individualistic and middle-class-oriented ideas of mental health. The challenge of algorithmic bias becomes especially serious in the fact that digital phenotyping systems – when they are trained on non-representative datasets, which for example mostly show white, middle-class populations – can reproduce systematic disadvantages that affect already marginalised youth groups.^63,64 This phenomenon, called ‘distributional shift’, shows up when machine learning models that were trained in limited populations (like university students) have unacceptable biases when used in real-life application contexts. ¹ In this connection, it must be considered that the recognition accuracy for culture-specific expressions of psychological stress – for example in adolescents with migration background or from non-Western cultures – can be significantly lower, which makes existing barriers to appropriate psychosocial care even worse.

These questions make clear that the explainability of digital phenotypes cannot be reduced to individualistic and procedural approaches, but must include structural justice dimensions, which combine distributive, representative and recognition-based aspects. Intersectional explainability is especially relevant in relation to marginalised adolescents, such as LGBTQ + persons or young people with migration or refugee background. The specific life situations, stress factors and resilience strategies of these groups may be pathologised or made invisible through standardisation based on algorithms. Justice-oriented explainability functions in this context as a critical resource to identify such mechanisms of exclusion, to problematise them and to develop transformative alternatives. The transformative explainability of digital phenotyping systems requires structural approaches, which go beyond technical optimisation or procedural adjustments and aim at fundamental changes in the power relations and institutional frameworks. A first step consists in the explicit recognition of the normative and political dimensions of digital phenotyping, which often stay hidden behind seemingly neutral scientific or medical discourses and must be made visible through critically-reflexive explainability. ⁴⁷ Participatory explainability enables the substantial involvement of diverse young people and marginalised groups in all phases of technology development, implementation and evaluation. In this way, a counterbalance to the dominant expert culture is created, and alternative perspectives on mental health and technological innovation can be integrated. ⁶² Structural explainability focuses on institutional frameworks and regulatory mechanisms which shape the use of digital phenotypes and can reproduce or transform social inequalities. The possibility of reflexive explainability finally enables a continuous critical analysis of the normative basic assumptions and value orientations that are inscribed into digital phenotyping systems, as well as a transparent societal debate about the adequate balance between technological innovation, social justice and youth autonomy. This multidimensional concept of transformative explainability underlines that the ethical design of digital phenotyping is a task for the whole society, which requires both individual and collective responsibility and presupposes continuous critical reflection and democratic deliberation.

The prioritisation of explainability against other ethical considerations like efficiency, accuracy or reach of care needs a more differentiated normative justification, that goes beyond just saying it is important. While some researchers, who highlight the opportunities of digital phenotyping,^1,3,8,65 rightly point out that other moral considerations – like the possibility to give more youths access to mental health care or the improvement of diagnostic precision – also represent strong ethical claims, one can argue for the special position of explainability in the context of youth mental health. This special role does not come from some abstract hierarchy of ethical principles, but from the constitutive role that explainability plays for realising other ethical goods. Without a minimal level of age-appropriate explainability, neither authentic autonomy development nor trustful therapeutic relationships can arise – both are conditions which themselves are prerequisites for the effectiveness of digital interventions. The importance of explainability does not mean that technical transparency must be maximised at any price. Rather, it is about establishing a threshold of context-sensitive explainability, under which the use of digital phenotyping with adolescents would violate the basic conditions of ethically acceptable healthcare. This threshold is not static, but must be defined depending on factors like the development status of the youths, the type of psychological strain, and the available alternatives. The position presented here is therefore not an absolutistic one, but one that understands explainability as a necessary, even if not sufficient, condition for the ethical use of digital phenotyping in adolescents – a condition whose concrete design must be subject of continuous negotiation between all involved actors.

Normative requirements for digital phenotyping in adolescents

Explainability as a normative precondition for the use of digital phenotypes in young people with psychological burdens becomes concrete in specific requirements for the technological, organisational and communicative design of this innovative form of diagnostics and intervention. The recognition of the special normative position of adolescents is here of fundamental importance. Young people move in a field of tension between childlike need for protection and adult autonomy. Their specific developmental tasks and potentials should not be negatively influenced by the use of digital technologies, but rather be supported. ²³ The fundamental requirement consists in the development of explainable AI systems that can make their function, decision logic and results transparent in a way that is understandable for different target groups – young people, parents, clinical staff. This technological explainability requires innovative approaches, which combine transparency with performance and translate complex algorithmic processes into comprehensible explanation patterns. At the same time, explainability must go beyond the purely technical dimension and also reveal the normative assumptions, epistemic uncertainties and potential bias that are written into the systems. The multidimensional transparency of digital phenotyping thus includes not only the disclosure of data collection and processing, but also the explication of the implemented values, norms and classification logics, as well as the limits and uncertainties of algorithmic predictions. This comprehensive transparency forms the basis for informed decisions and the continuous critical reflection and evaluation of digital phenotyping practices. The development-appropriate design of digital phenotyping requires a differentiated consideration of the cognitive, emotional, and social developmental dynamics of adolescence. The information transfer and consent processes must be adapted to the specific understanding abilities and needs of adolescents. For this, the use of age-specific communication forms and channels is necessary. In addition, complex technological and medical concepts must be translated into everyday-relevant, understandable explanations. ³⁹ The emotional dimension plays here a decisive role, because the diagnosis and monitoring of psychological burdens can be connected with shame, fear or experiences of stigmatisation, which influence the information processing and decision-making.

The social embedding of technology use must reflect the complex relationship dynamics between adolescents, parents, and therapists, ⁶⁶ whereby the technology should not undermine existing relationships and processes of autonomy development, but rather support them. Furthermore, the temporal dimension of development must be considered, by establishing continuous information and consent processes adapted to the developmental status, which respect and foster the growing autonomy capacities of adolescents. This development-sensitive design requires a deep understanding of adolescence beyond simplifying categorisations and stereotypical attributions, and should be based on developmental psychological and educational findings, which are complemented and validated through participatory research with adolescents.^38,67 The institutional anchoring of explainability as a normative precondition requires the establishment of robust governance structures that go beyond individual education and consent practices and create structural conditions for responsible digital phenotyping. This includes the development of specific ethical guidelines and standards for the explainability of AI systems in adolescent mental health care. These guidelines and standards address technical as well as communicative and social aspects and define concrete minimum requirements. The integration of these standards into approval, certification, and evaluation processes can increase their binding nature and create incentives for the development of explainable systems.

There is an emphasis on the need for independent oversight and evaluation mechanisms to monitor the compliance with these standards and to intervene in case of deviation. In this process, not only experts, but also representatives of the target groups – adolescents, parents, and patient organisations – should be included. The continuous training and sensitisation of all involved actors, especially of clinical personnel, is of crucial importance in order to establish explainability not only as a technical, but also as a communicative and ethical competence. In addition, specific support structures for adolescents and parents are required, which provide assistance in navigating complex information and decisions. Among these structures are specialised counselling services, peer-support programmes or digital educational resources. The societal responsibility for the ethical design of digital phenotyping finally demands a broad public debate about the adequate balance between innovation, autonomy, and protection in the context of adolescent mental health. This debate must integrate different perspectives and values and be reflected continuously.

Practical implications for the implementation of explainable digital phenotypes

The implementation of the developed normative requirements for explainable digital phenotyping in clinical everyday practice is confronted with a variety of practical challenges, while at the same time innovative solution strategies are needed. One central practical implication concerns the design of human-technology interaction, which includes both the technological architecture and the user experience. Especially promising appear adaptive explanation systems that can offer different levels and formats of explanation and adapt them to the individual context, level of knowledge, and the specific information needs of the users. ³⁵

Adaptivity can be supported by interactive elements that allow adolescents to actively ask questions, give feedback and express their preferences concerning type and extent of explanations. The visual and narrative design of the explanations should use youth-appropriate aesthetics, language and metaphors, without becoming simplifying or infantilising. Furthermore, the integration of the explanation systems into existing clinical workflows has to be taken into account. Here, both the time and cognitive resources of the medical staff and the specific requirements of different use contexts – such as initial consultation, continuous application, or crisis intervention – must be considered. The technical design must always be guided by ethical and pedagogical considerations, which place the specific needs and developmental dynamics of adolescents in the centre and put the technology in service of therapeutic relationships and adolescent autonomy development.

The practical implementation of explainable digital phenotypes requires innovative educational and empowerment strategies, which go beyond traditional information conversations and aim at the promotion of technological and health-related competences among all involved persons. The development of a basic AI literacy for adolescents, as suggested by Ng et al., ⁵⁸ is one important building block, but should be integrated into a broader promotion of health and media literacy. For this purpose, age-appropriate educational formats are suitable, which can make use of both formal and informal learning contexts and integrate digital, interactive, and playful elements. The implementation of peer education approaches appears as particularly promising, as they make use of specific communication and learning forms of adolescents and at the same time can promote their autonomy and self-efficacy.^29,58

At the same time, there is a need for specific training for parents, which enables them to communicate complex technological and medical concepts in an age-appropriate way and to act as trustworthy interpretation instances. For clinical professionals, interdisciplinary training offers should be developed, which convey both basic technological knowledge and communicative and ethical competences and show practice-oriented strategies for the integration of digital phenotypes into therapeutic relationships and treatment concepts. The effectiveness of these educational measures should be evaluated continuously and adapted to new technological developments and experiences. In this, the inclusion of all stakeholders, especially the adolescents themselves, is of central importance.^68-70 The integration of relational and procedural approaches into clinical practice requires the development of innovative organisational structures and processes, which consider both the normative requirements for explainability and informed consent, and the practical conditions of clinical routine. Instead of standardised consent procedures, the establishment of multi-step, dialogical processes is recommended, which include both formal and informal elements and enable continuous reflection and adaptation. Innovative documentation forms can support the making visible of the procedural and relational aspects of informed consent and at the same time fulfil legal requirements.

The implementation of multidisciplinary teams, which include not only medical professionals but also expertise in the fields of technology, ethics, pedagogy and communication, can support holistic accompaniment of adolescents and their families and bring in various perspectives on the chances and risks of digital phenotyping. It is essential to integrate specific time windows and spaces for explanation, reflection and decision-making into the clinical workflow, to do justice to the complexity of the topic and the needs of the involved persons. In addition, the definition of clear responsibilities and escalation paths for ethical dilemmas or technical problems that may occur in the context of digital phenotyping is essential. The continuous evaluation and adaptation of these organisational structures and processes in dialogue with all participants forms an essential precondition for the ethically acceptable and practically feasible integration of digital phenotypes into adolescent mental health care.

Conclusion and outlook: Explainability as ethical foundation of digital transformation

Digital phenotyping represents a paradigmatic example for the profound transformation of healthcare through digital technologies, whereby the special vulnerability and developmental dynamics of adolescents and the sensitivity of mental health create specific ethical requirements. This analysis has identified explainability as a fundamental normative precondition for the ethically responsible use of digital phenotypes in adolescents with psychological burdens and elaborated it in its various dimensions. Explainability turns out not to be a purely technical or procedural concept, but a multidimensional normative construct, which includes technological, epistemic, clinical, and ethical aspects and must be realised in complex social and institutional contexts. The relational perspective on adolescent autonomy and informed consent extends the traditional bioethical discourse and allows a more differentiated understanding of developmental dynamics and interpersonal relationships that shape the context of digital phenotyping. Especially the integration of trust, transparency and critical reflection of power dynamics into the concept of explainable digital phenotypes opens up new perspectives on the ethical design of digital health technologies, which go beyond individual consent questions and address structural dimensions of justice and inclusion. The normative requirements that have been developed for explainable digital phenotyping offer a guiding framework for how this can be implemented in clinical practice, but finding the right balance between ethical ideals and what is practically doable stays a constant challenge that needs further interdisciplinary research and evaluation close to real-life practice. The future development of explainable digital phenotypes will be strongly influenced by technological innovations, changes in social values, and the legal and regulatory environment. From the technological side, promising progress can already be seen in areas like explainable AI, federated learning, and privacy-preserving computing, which could offer new ways to balance explainability, data protection, and system performance.

At the same time, the societal debate on the appropriate role of digital technologies in adolescent mental health care will become more intense, where questions of medicalisation, digital surveillance and the relation between technological and human care will come into focus. The regulatory landscape will continue to evolve, with increasingly specific requirements for AI in healthcare, whereby the balance between promotion of innovation and protection of vulnerable groups becomes a central challenge. For future research and practice, many tasks arise: the development and evaluation of concrete methods and tools for explainable digital phenotypes; the empirical investigation of actual impacts of digital phenotyping on adolescent autonomy development and therapeutic relationships; the creation of specific training concepts for all involved persons; and the continuous ethical reflection of new technological developments and application areas.

Particular importance belongs to participatory research and development approaches, which do not see adolescents only as research objects or technology users, but include them as active co-designers throughout the whole process and take their perspectives, needs and values seriously. Only through this continuous reflection and participatory design can the transformative potential of digital phenotypes for adolescent mental health be realised in an ethically responsible way, without compromising fundamental values like autonomy, privacy, and social justice.