Ecological momentary intervention

1. Introduction

Version followed: the Heron and Smyth (2010) definition of ecological momentary intervention (EMI) as summarised by Balaskas et al. (2021); ACT in Daily Life (ACT-DL) as delivered in the INTERACT trial (Myin-Germeys et al., 2021 medRxiv preprint); the Therap-i personalised experience-sampling feedback module (von Klipstein et al., 2023); the just-in-time adaptive intervention (JITAI) component framework of Nahum-Shani et al. (2018) as set out by Walton et al. (2020 draft); and the m-Path platform as described by its developers (Mestdagh et al., 2023).

What it is. EMI is a delivery format, not a single therapy. It means delivering treatment components in everyday life, in real time and in the person's natural settings, usually by smartphone, either standalone or as a supplement to existing treatment [source: Balaskas et al. 2021, citing Heron & Smyth 2010]. It is closely tied to ecological momentary assessment (EMA), also called the experience sampling method (ESM): repeated brief self-reports of mood, symptoms, thoughts, activities and context, collected in daily life as they happen rather than recalled later. Momentary and retrospective reports can diverge substantially; in one early study, coping recorded close to a stressor matched poorly with coping recalled within 48 hours [source: Stone et al. 1998]. EMA data can then decide the timing and content of the intervention, sometimes labelled EMA-informed EMI [source: Balaskas et al. 2021]. JITAIs are a subset of EMI that adapt the type, amount and timing of support to the person's changing internal state and context, using predefined decision rules [source: Balaskas et al. 2021; Nahum-Shani et al. 2018 abstract].

How change is thought to happen. Three mechanisms are described. (1) Transfer: skills learned in session are practised at the moments they are needed, so ACT-DL pairs each face-to-face session with daily-life prompts carrying an exercise or metaphor from that session [source: Myin-Germeys et al. 2021]. (2) Insight from one's own data: client and therapist test their own case hypotheses against daily-life data and build the formulation together [source: von Klipstein et al. 2023]. (3) Timeliness: support given when the person is vulnerable and able to engage is expected to have more effect than support at arbitrary times [source: Nahum-Shani et al. 2018 abstract]. For a psychotherapist, EMI is used mainly as between-session support and monitoring within an ongoing therapy [source: Balaskas et al. 2021; Mestdagh et al. 2023].

Research versus clinical status (as of October 2026, per the sources listed; the most recent is dated 2023). EMI remains mainly a research-stage method. A 2021 scoping review found that clinical validation was at an early stage: 9 RCTs among the included studies [source: Balaskas et al. 2021]. In the multicentre INTERACT trial of ACT-DL in early psychosis, ACT-DL plus treatment as usual did not beat treatment as usual on the primary interview-rated outcome of psychotic distress. It did reduce momentary psychotic distress measured by ESM, and it improved negative symptoms and global functioning; these results come from a preprint not yet peer reviewed [source: Myin-Germeys et al. 2021]. Whether a peer-reviewed version changed these results is [uncertain: not verified]. The m-Path developers described medical-device certification of their platform as under investigation in 2023 [source: Mestdagh et al. 2023]. No FDA clearance, DiGA listing, NICE recommendation or CE/UKCA medical-device marking specific to EMI as a category was identified [uncertain: not verified]. Check the status of any specific product yourself before use, because product status changes. A general-purpose ESM platform used to send self-help content is not, by that fact, an approved treatment device [synthesis].

Who uses it. Psychotherapists, psychologists and other mental-health clinicians can design and run EMI within their own scope of practice, using clinician dashboards [source: Mestdagh et al. 2023; Balaskas et al. 2021]. Sensor-driven and machine-learning JITAIs remain research tools [source: Balaskas et al. 2021].

2. Framework

EMI has no single phase model. The phases below combine the procedures described in ACT-DL, Therap-i and the platform literature; the sequencing is [synthesis].

Phase 0: Suitability, consent and set-up (1 session or part of one).

• Goals: decide whether momentary data or prompts would add something the therapy needs, and agree targets.
• Goals: obtain informed consent covering burden, data handling and monitoring limits.
• Goals: install and test the app.
• Move on when: the client understands what is collected and who sees it, and knows the app is not monitored around the clock unless you have arranged that. A crisis plan exists. A trial prompt has been completed together [synthesis; source for the monitoring-limits element: Bentley et al. 2021].

Phase 1: Assessment or formulation sampling (typically 1–2 weeks, or longer).

• Goals: collect personalised EMA data to test hypotheses about triggers, contexts and patterns.
• In Therap-i the client and therapist co-designed the questionnaire. The client answered it five times a day for eight weeks, alongside usual psychotherapy [source: von Klipstein et al. 2023].
• Move on when: there is enough data to look for patterns. Compliance is acceptable, or the barriers to it have been addressed [synthesis].

Phase 2: Collaborative feedback (one or more sessions).

• Goals: review graphs together, let the client interpret them, and refine the formulation and treatment targets.
• Therap-i used three feedback sessions with patient, therapist and a researcher [source: von Klipstein et al. 2023]. In the trial protocol, the personalised diary is built jointly by patient, clinician and researcher from the patient's case conceptualisation, and feedback is given three times during the eight-week monitoring period and discussed by all three [source: Riese et al. 2021 protocol].

Phase 3: Intervention in daily life.

• Goals: deliver brief, session-linked exercises in daily life, as prompted (push) and on-demand (pull) content.
• ACT-DL model: 8 manualised face-to-face sessions of 45–60 minutes. Session 1 is psychoeducation, sessions 2–7 each cover one of six ACT components, and session 8 integrates them. After each session from session 2 on, the app sends 8 semi-random prompts a day for 3 days. Each prompt has a brief momentary questionnaire and an exercise or metaphor from the preceding session, and exercises are also available on demand [source: Myin-Germeys et al. 2021].
• Move on when: the client can use the skill without prompting, or the proximal target is changing [synthesis].

Phase 4: Monitoring, tapering and ending.

• Goals: optionally continue light EMA as outcome monitoring. Then reduce prompts, agree what happens to the data, and end app access.
• In ACT-DL, app access ended with the intervention period [source: Myin-Germeys et al. 2021]. m-Path lets participants delete their data or block access [source: Mestdagh et al. 2023].

Format. Individual therapy with an app. Clinicians build questionnaires and schedules in a web dashboard and view each client's data as graphs [source: Mestdagh et al. 2023]. Clinicians configure EMIs manually in many studies; others use semi-individualised selection from a common set, and fully automated tailoring is rare [source: Balaskas et al. 2021].

Age limits. The cited sources set no general age limit for EMI [uncertain]. The INTERACT trial included people aged 15–65 [source: Myin-Germeys et al. 2021]. For minors, experts strongly agreed that a person should reach out after a high-risk response, and parental contact raises extra issues [source: Bentley et al. 2021]. Check the app's own terms of use for minimum ages and local rules on minors' consent to data processing [synthesis].

3. Guiding principles

• EMI supplements therapy; it does not replace it. Use the app to carry session work into daily life and bring daily-life data back into the session [source: Balaskas et al. 2021; Myin-Germeys et al. 2021].
• Start from a clinical question. Every item should test a hypothesis you and the client hold about the case [source: von Klipstein et al. 2023].
• Co-create, don't impose. Client and therapist design items together, and the client interprets their own data first [source: von Klipstein et al. 2023].
• Keep questionnaires short. Longer questionnaires raised burden and lowered data quantity and quality. Higher sampling frequency did not show those effects in students [source: Eisele et al. 2022]. Cut items before you cut prompts.
• Use the least push content that works. Frequent pushes irritate, cause disengagement and lose effect through habituation, so keep the components that earn the interruption [source: Walton et al. 2020].
• "Do nothing" is a valid option. At a decision point, not prompting can be the right choice, for example when the person is unavailable, such as while driving [source: Walton et al. 2020].
• Link the micro-intervention to the last session. The prompt rehearses what was just taught, so the client recognises it [source: Myin-Germeys et al. 2021].
• No clinical decisions from unvalidated signals. Sensor-based and machine-learning predictions of stress or relapse are still in development [source: Balaskas et al. 2021]. Use them, if at all, only as prompts for a conversation, never to diagnose, change risk level or alter treatment [synthesis].
• A momentary suicidal-ideation report shows that the thought occurred. It does not on its own show that behaviour is imminent [source: Bentley et al. 2021]. Respond with a clinical risk assessment, not an automatic escalation [synthesis].
• Be honest about monitoring limits. Tell clients plainly whether, when and by whom their entries are seen [source: Bentley et al. 2021].
• Collect the minimum data needed. Prefer platforms that need no name or phone number, with encryption and data deletion rights [source: Mestdagh et al. 2023].
• Offer an alternative. When a client is worried about electronic data privacy, a paper version of the exercises is an option [source: Balaskas et al. 2021].
• The clinician stays responsible. You choose content, review data, handle risk and decide on treatment. The software only delivers and records [synthesis].

4. Intervention inventory

The procedures below cover the full workflow, from deciding whether to use EMI to ending. Platform functions are cited to Mestdagh et al. (2023). Clinical steps are cited where a source describes them and marked [synthesis] where they are general practice. Prompt and exercise design is described in general terms only. ACT-DL's manualised session content and exercises are not reproduced here; obtain the manual from its developers [uncertain: not verified whether it is publicly available].

4.1 Suitability check and choosing whether to use EMA, EMI or both

• Purpose / when: At the start of therapy, or when progress stalls because in-session recall of daily life is vague or contradictory. Momentary reports can differ from what is recalled later [source: Stone et al. 1998].
• Steps:
  1. Name the clinical question. It might be about the formulation ("what precedes my binge urges?"), about skill transfer ("I forget to use defusion when it matters"), or about monitoring ("is my mood actually improving?").
  2. Pick the mode:
     • EMA only, for formulation or monitoring [source: von Klipstein et al. 2023].
     • EMI, for prompted practice [source: Myin-Germeys et al. 2021].
     • Both, with EMA items at each prompt plus a linked exercise, as in ACT-DL.
  3. Check the practicalities: smartphone access, reading level, the device model (some recent Huawei phones are not supported by m-Path) and internet access, since notifications need a connection [source: Mestdagh et al. 2023].
  4. Screen for reasons to delay or adapt [synthesis]:
     • current high risk without a working crisis plan;
     • persecutory ideas about being monitored;
     • an eating disorder or health anxiety in which self-monitoring may feed checking;
     • a coercive partner who could read the phone.
  5. Decide who will see the data, and when [source for monitoring-limits element: Bentley et al. 2021].
• Example language (illustrative): "When we talk about last week, we're both relying on memory, and memory smooths things over. Would it help to catch a few moments as they happen, with short check-ins on your phone, and look at them together?"
• Watch for: A client who agrees in order to please you. Paranoid or shame-laden reactions to the idea of being tracked. A device shared with others. It is working if the client can say what question the data will answer [synthesis].

4.2 Informed consent and data-privacy agreement

• Purpose / when: Before any data collection. Separate from consent to therapy [synthesis].
• Steps:
  1. Explain what is collected: questions only, or also sensor data such as location or phone use, which some EMIs use [source: Balaskas et al. 2021]. Explain how often and for how long.
  2. Explain who stores the data, and where. For example, m-Path stores data on servers in Belgium, states that it is GDPR compliant, encrypts phone-server traffic and needs no name, email or phone number [source: Mestdagh et al. 2023]. For any other platform, check the equivalent facts yourself, along with HIPAA, GDPR or local equivalents [synthesis].
  3. State the monitoring limits plainly: whether entries are reviewed in real time, daily, or only at the next session. Include what you will do if a risk item is endorsed [source: Bentley et al. 2021]. Studies varied widely here, and some told participants at consent that data are not monitored in real time [source: Bentley et al. 2021].
  4. Explain the client's rights. They can stop at any time, delete their data, block your access or choose to share it [source: Mestdagh et al. 2023]. A paper alternative is available [source: Balaskas et al. 2021].
  5. Agree how the data will be recorded in the clinical file. Exported graphs can be added to records [source: Mestdagh et al. 2023]. Agree what happens to the data at the end.
  6. Record the consent. An intake interaction in the app can carry a consent form [source: Mestdagh et al. 2023], but discuss consent face to face as well [synthesis].
  7. For minors: involve the guardian as local law requires, and agree in advance who is contacted after a high-risk response [source: Bentley et al. 2021; synthesis].
• Example language (illustrative): "I'll be honest about what this app can and can't do. I look at your answers before each session, not as they come in. If you're in crisis between sessions, the app won't reach me. That's what your crisis plan is for. We'll put the numbers in the app too."
• Watch for: A client who assumes you are watching live. Ask them to say the monitoring arrangement back to you. Watch also for hesitancy about recording sensitive behaviour such as substance use or self-harm; offer to omit those items or use paper [synthesis].

4.3 Choosing targets and co-creating items

• Purpose / when: To turn the formulation into a few momentary questions.
• Steps:
  1. From your current formulation, list two or three working hypotheses together. In the Therap-i case these concerned low energy, explored through sleep, medication and activity, and low self-esteem, explored through thinking patterns and social interactions [source: von Klipstein et al. 2023].
  2. For each hypothesis, choose an outcome item (such as momentary mood or urge), a proposed driver (a thought, an activity, company) and a context item (where and with whom) [synthesis, consistent with von Klipstein et al. 2023].
  3. Word the items in the client's own language. In Therap-i, personalised items included self-esteem and open versus withdrawn behaviour [source: von Klipstein et al. 2023].
  4. Choose formats: sliders or Likert scales for intensity, yes/no for events, multiple choice for context, and free text sparingly. Use branching, so that follow-up questions appear only when relevant [source: Mestdagh et al. 2023].
  5. Keep the total short. Long questionnaires increased burden and reduced data quality [source: Eisele et al. 2022]. Validated item banks exist; use established items where they fit rather than inventing scales [uncertain: no specific item repository is recommended here].
  6. Avoid putting standardised copyrighted questionnaires into the app without a licence [synthesis].
• Example language (illustrative): "You said you feel worst after you've been alone all afternoon. Let's test that. One question about how low you feel right now, one about whether you've been with anyone in the last hour. We'll see what the data says, and it may surprise both of us."
• Watch for: Too many items. Items that only you care about. Items that read as judgments, such as "Did you avoid again?" Signs it is working: the client proposes items, and can predict what the graph might show [synthesis].

4.4 Designing the sampling schedule

• Purpose / when: To decide when prompts arrive, so the data fit the question and the client's day.
• Steps:
  1. Choose a schedule type [source: Balaskas et al. 2021; Mestdagh et al. 2023]:
     • fixed times;
     • random or semi-random times within windows;
     • event-contingent, where the client taps a button when the target event happens (an urge, a conflict, a voice);
     • user-initiated;
     • a combination.
  2. Set the prompt frequency, the response window before a prompt expires, and any reminder [source: Mestdagh et al. 2023]. Published examples include 5 prompts a day for 8 weeks in Therap-i and 8 semi-random prompts a day for 3 days after each session in ACT-DL [source: von Klipstein et al. 2023; Myin-Germeys et al. 2021]. Fit the frequency to the client; higher frequency was not linked to extra burden in students [source: Eisele et al. 2022], but clinical groups may differ [synthesis].
  3. Block out times when the client must not be interrupted: work meetings, driving, sleep. These are "unavailable" contexts where no prompt should be sent [source: Walton et al. 2020].
  4. Set the start and end dates, and plan for gaps such as holidays [synthesis].
  5. Test the schedule on your own phone first. m-Path offers a self-test function [source: Mestdagh et al. 2023].
• Example language (illustrative): "Random times catch ordinary moments. A button you press when the urge hits catches the moments we most want to understand. Shall we do both, and switch the prompts off during your shifts?"
• Watch for: Prompts during unsafe moments such as driving. Clusters of missed prompts at the same time of day, which suggests the schedule needs changing rather than the client needing persuading. It is working if response rates are stable [synthesis].

4.5 Designing momentary micro-interventions (prompt content, in general terms)

• Purpose / when: To deliver brief practice of a skill already introduced in session, at a moment in daily life.
• Steps:
  1. Select one skill from the most recent session. In ACT-DL, each prompt offered an exercise or metaphor tied to the component covered in the previous session [source: Myin-Germeys et al. 2021].
  2. Choose a format the client will actually use [source: Balaskas et al. 2021; Mestdagh et al. 2023]:
     • one or two lines of text;
     • a short audio guide recorded in your voice or the client's;
     • an image or a short video;
     • a link to an on-demand library.
  3. Make it doable in under a minute or two, wherever the client is [synthesis].
  4. Option: place a short EMA item before it, so that what is shown depends on the answer. Branching rules based on EMA responses were the commonest way of individualising feedback [source: Balaskas et al. 2021].
  5. Avoid instructions that are unsafe out of session. Exposure tasks, trauma processing or anything likely to cause marked distress without support should not be prompted remotely unless your protocol includes safety steps and the client has consented [synthesis].
  6. Give clinical content in your own words or from licensed materials [synthesis].
• Example language (illustrative), describing a prompt in session: "When your phone buzzes, there'll be one question about how stuck you feel, then a ten-second reminder of the 'passengers on the bus' idea we used today. You don't have to do anything else."
  • The metaphor named is a common ACT metaphor, used here as an example; whether it appears in ACT-DL specifically is [uncertain: not verified].
• Watch for: Content that feels generic or nagging. Skills that have not yet been learned in session. Exercises the client experiences as self-criticism. It is working if the client reports using the skill in moments without being prompted [synthesis].

4.6 Choosing the tailoring level (manual, semi-individualised, automated) and JITAI rules

• Purpose / when: To decide how much the system adapts by itself.
• Steps:
  1. Default to manual tailoring: you and the client configure content and schedule and revise them at each session. Clinicians tailored in this way in a number of studies, mainly through portals [source: Balaskas et al. 2021].
  2. Semi-individualised: the platform selects from a set you approved, based on the client's preferences or your choices [source: Balaskas et al. 2021].
  3. A simple JITAI, if your platform supports conditional rules. Define each element in writing [source: Walton et al. 2020; Nahum-Shani et al. 2018]:
     • the distal outcome (the treatment goal);
     • the proximal outcome the prompt should shift (for example, urge intensity within the next hour);
     • the decision points;
     • the intervention options, including doing nothing;
     • the tailoring variables (for example, a self-reported stress rating above the client's agreed threshold);
     • the decision rule linking them.
  4. Use only self-report tailoring variables the client understands. Do not let unvalidated sensor or machine-learning estimates drive content in clinical use [source for development status: Balaskas et al. 2021; rule: synthesis].
  5. Review each rule with the client. Remove components that do not shift the proximal outcome enough to justify the interruption [source: Walton et al. 2020].
• Example language (illustrative): "We can set it so that if you rate your stress above 7, it offers the breathing audio. If you're below that, it just says thanks. Does 7 feel like the right line?"
• Watch for: Rules firing too often. A client who games ratings to avoid content. The illusion that the app "knows" the client. It is working if prompts arrive at moments the client agrees were useful [synthesis].

4.7 Onboarding and practice run

• Purpose / when: At the end of the set-up session.
• Steps:
  1. The client installs the app in session. In m-Path they choose an alias, save a recovery code and connect using your name or an invitation code [source: Mestdagh et al. 2023].
  2. Complete one prompt together, including an event button if used.
  3. Check the phone's notification settings and battery optimisation, which can stop background apps [source: Mestdagh et al. 2023].
  4. Agree what to do if the app fails: note it on paper and tell you at the next session [synthesis].
  5. Put crisis contacts inside the app as well as on paper. Some studies made safety plans available in the app [source: Bentley et al. 2021].
• Example language (illustrative): "Let's do the first one now so nothing's a mystery. Read it out and tell me if any question feels odd or hard to answer."
• Watch for: Confusion over items. A phone that is shared or monitored by others. Disabled notifications [synthesis].

4.8 On-demand (pull) exercise library

• Purpose / when: To give the client self-initiated access to session skills alongside prompted (push) content.
• Steps:
  1. Add short exercises from completed sessions to an on-demand menu. ACT-DL participants could open exercises whenever they needed them [source: Myin-Germeys et al. 2021]. m-Path supports self-help pages with buttons [source: Mestdagh et al. 2023].
  2. Keep the menu small and named in the client's words [synthesis].
  3. Ask at each session which items were used, and in what situations [synthesis].
• Example language (illustrative): "The prompts come to you. This menu is for when you notice you need something. Which two exercises do you want there this week?"
• Watch for: Pull content relies on the client noticing their own need. Push content does not, but it interrupts [source: Walton et al. 2020]. A client who never opens the menu may need more push content, or a simpler menu [synthesis].

4.9 Integrating EMI with sessions (session-linked prompting blocks)

• Purpose / when: To structure therapy so each session seeds daily-life practice and the next session harvests it.
• Steps:
  1. Teach the component in session. At the end, agree on that week's prompt content [source: Myin-Germeys et al. 2021].
  2. Run a short, intensive prompting block after the session. ACT-DL ran 3 days of prompts after each session [source: Myin-Germeys et al. 2021].
  3. Open the next session by reviewing what happened with the prompts and exercises (see 4.10) [synthesis].
  4. Adjust the content or the skill accordingly, then introduce the next component [synthesis].
  5. Where content from sessions can be loaded into the app, do so, so that the client sees continuity [source: Balaskas et al. 2021].
• Example language (illustrative): "For the next three days, the check-ins will carry today's 'noticing' exercise. Next week we'll start by looking at when it helped and when it didn't."
• Watch for: Session content drifting away from the app content. Prompts continuing into material not yet covered. Attendance and app use moving together; in ACT-DL, participants attended an average of 6 of 8 sessions [source: Myin-Germeys et al. 2021].

4.10 Reviewing data with the client (collaborative feedback session)

• Purpose / when: After a sampling period (for formulation) and periodically during EMI (for progress). This is the step where EMA becomes therapy.
• Steps:
  1. Before the session, look over the data yourself in the dashboard: time-series graphs, frequencies, weekly overviews and comparisons between time ranges [source: Mestdagh et al. 2023]. Note the compliance rate and any risk entries first [synthesis].
  2. In session, show the graphs on a shared screen or the client's own app view. Clients can see their own responses over time [source: Mestdagh et al. 2023].
  3. Ask the client to interpret first, then add your observations. Therap-i used a flexible visual interface that patient and therapist explored together, with the patient actively interpreting [source: von Klipstein et al. 2023].
  4. Go back to the hypotheses from 4.3. Were they supported, contradicted, or unclear? [source: von Klipstein et al. 2023].
  5. Treat patterns as hypotheses, not facts. Missing data, few data points and careless responding all limit what you can conclude [source for careless responding: Eisele et al. 2022; interpretation: synthesis].
  6. Translate insights into formulation and plan changes. In the Therap-i case, feedback led to insights about energy and self-esteem that informed treatment [source: von Klipstein et al. 2023].
  7. Save the agreed summary or exported graph to the record [source: Mestdagh et al. 2023].
• Example language (illustrative):
  • "Before I say anything, what jumps out at you?"
  • "So on the days you saw your sister, your mood ratings sit higher. What do you make of that?"
  • "This is two weeks of data with gaps, so let's treat it as a lead rather than proof."
• Watch for:
  • The client feeling judged by the data ("I failed again"), or the therapist over-interpreting.
  • Graphs that contradict the client's account. Handle these with curiosity, not confrontation (see 5).
  • Signs it is working: the client refers back to their data spontaneously, and the formulation becomes more specific [synthesis].

4.11 Ongoing momentary outcome monitoring

• Purpose / when: To track change between sessions as part of routine outcome monitoring [source: Mestdagh et al. 2023]. In several studies, therapists used EMI data running in parallel with face-to-face sessions to track progress [source: Balaskas et al. 2021].
• Steps:
  1. Keep two to four core items constant across therapy, so that they can be compared over time [synthesis].
  2. Use a light schedule, for example a few prompts on selected days [synthesis].
  3. Review trends at agreed intervals, and alongside your standard retrospective measures, which momentary data do not replace [synthesis].
  4. Discuss divergence between momentary and retrospective measures as clinical information. Momentary and recalled reports can differ [source: Stone et al. 1998].
• Example language (illustrative): "Your monthly questionnaire says things are much the same, but your daily ratings show more good afternoons than in March. Which fits your experience?"
• Watch for: Monitoring fatigue, and "flat" responding that suggests answers given by rote [source on careless responding: Eisele et al. 2022]. Drop or pause monitoring if its burden outweighs its use [synthesis].

4.12 Safety monitoring and crisis handling of in-the-moment reports

• Purpose / when: Whenever any item touches on suicidal thoughts, self-harm, violence, substance use or acute symptoms. Also when free text could carry a disclosure. Settle this before sampling starts. Even with no risk items, a client can type alarming free text, so set a policy for every EMA [synthesis].
• Steps:
  1. Decide whether risk items are needed at all. Do not ask about suicidality in an app unless you can respond to the answers [synthesis]. Evidence so far suggests that asking frequently about suicidal thoughts in EMA does not increase them [source: Bentley et al. 2021].
  2. Define thresholds in advance. Studies used triggers such as any current ideation, ideation severity, intent, plan, a recent attempt or recent self-harm, and thresholds varied widely [source: Bentley et al. 2021]. Agree yours with your team or supervisor [synthesis].
  3. Build automatic in-app responses for threshold entries. The commonest content in studies was encouragement to call a crisis line, contact existing supports or use emergency services. Fewer studies included the study team's contact details or a safety plan [source: Bentley et al. 2021]. Use the client's own safety plan and local crisis numbers, and say plainly in the message whether the entry will be seen soon [source: Bentley et al. 2021; synthesis].
  4. Set and keep a review rhythm you can actually sustain. In studies, review ranged from real-time alerts to daily or less often. An expert panel agreed with about 90% consensus that data should be reviewed at least every weekday, or in real time where the platform allows [source: Bentley et al. 2021]. Not every platform sends risk alerts; m-Path's developers do not describe a crisis-alert feature [source: Mestdagh et al. 2023]. If you cannot review promptly, do not include risk items, and tell the client [synthesis].
  5. Respond to high-risk entries with direct contact. In 94.7% of monitoring studies, staff contacted the participant directly, usually by phone, for a suicide risk assessment. A clinician or senior staff member made the call, and other treating providers were sometimes notified [source: Bentley et al. 2021]. In practice, follow your service's risk-assessment and escalation policy, and document the entry, the contact and the decision [synthesis].
  6. Calibrate the response. A momentary report of suicidal thoughts does not on its own mean imminent behaviour [source: Bentley et al. 2021]. Assess in the usual way, covering ideation, intent, plan, access to means, protective factors and current safety, and review the safety plan, including lethal-means counselling (safe storage or removal). Avoid reflexive emergency escalation [synthesis].
  7. Respect the client's concerns about emergency services. Some research teams stopped listing emergency services because of concerns about police involvement, especially for Black and Hispanic participants [source: Bentley et al. 2021]. Discuss in advance which crisis options the client will use, and which ones you may have to use under your duty of care [synthesis].
  8. Minors. Experts strongly agreed that a person should reach out after a high-risk response from a minor. Agree in advance how and when parents will be involved, since parental contact raises its own concerns [source: Bentley et al. 2021].
  9. Lower-risk entries. Experts agreed only partly (about 61%) that an in-app pop-up alone is enough for low or moderate risk responses [source: Bentley et al. 2021]. Review these at the next contact at the latest [synthesis].
• Example language (illustrative):
  • At consent: "If you answer the thoughts-of-suicide question above 'a little', the app will show your safety plan and the crisis line straight away. I check entries each weekday morning. If I see a high one, I'll phone you that day. Outside those times, the app is not a way to reach me."
  • After an entry: "I saw that on Tuesday evening you rated the urge to hurt yourself quite high. I'm glad you were honest. Can we go through what was happening, and where things are now?"
• Watch for:
  • Under-reporting or stopping responses after an outreach call; the review raises this concern but has not tested it [source: Bentley et al. 2021].
  • Clients who believe they are watched live.
  • Entries landing at weekends or outside your hours.
  • Your own response drift, such as checking less over time.
  • It is working if risk entries lead to timely, documented contact and the client continues to report honestly [synthesis].

4.13 Treating sensor, passive and algorithmic signals cautiously

• Purpose / when: When a platform offers passive data (location, activity, phone use) or machine-learning estimates of stress, mood or relapse risk.
• Steps:
  1. Know the status of these tools. Sensing and machine-learning detection of stress, depression, schizophrenia and alcohol-related states have mostly been at the development stage [source: Balaskas et al. 2021].
  2. Do not base diagnosis, risk level, level of care, medication suggestions or other treatment decisions on such signals [synthesis].
  3. If a signal is displayed, treat it as a question to raise with the client ("Your step count dropped this week. How has getting out been?"), never as a finding [synthesis].
  4. Collect passive data only with explicit consent and a clear purpose. Turn it off when it has no clinical use [synthesis].
  5. Questions about medication raised by the data go to the prescriber. Never suggest starting, stopping or changing a medicine [synthesis].
• Example language (illustrative): "The app makes a guess about your stress from how you use your phone. Those guesses aren't reliable enough to act on, so I'd rather hear from you directly."
• Watch for: Clients or colleagues treating a dashboard score as a diagnosis. Automation bias in yourself. Location tracking with abuse or legal implications [synthesis].

4.14 The clinician's review, supervision and documentation routine

• Purpose / when: Throughout. EMI adds work outside sessions, so it needs scheduled time.
• Steps:
  1. Block a fixed time to review data that matches your consent statement (see 4.12) [source for review-frequency norms: Bentley et al. 2021].
  2. Check compliance, risk entries, technical failures, and whether the content still fits the session [synthesis].
  3. Bring EMI cases to clinical supervision, especially the risk protocol and any automated rules [synthesis].
  4. Keep a written protocol listing items, schedule, rules, thresholds, responders and review times. Record changes [synthesis].
  5. Clinician involvement is common in published EMIs. In studies it included remote support with technical problems and adherence, in-app messaging, and portals for monitoring adherence and sending tailored messages [source: Balaskas et al. 2021].
• Example language (illustrative), to the client: "I'll look at your check-ins on Monday and Thursday mornings. If something looks off with the app, I'll text you."
• Watch for: Unsustainable promises. Data piling up unreviewed. Messaging through the app turning into an unbounded on-call channel [synthesis].

4.15 Tapering, ending and data closure

• Purpose / when: At the end of an EMI block or of therapy.
• Steps:
  1. Taper the number of prompts, and shift towards pull-only use before ending, so the client relies on internal cues [synthesis].
  2. Review the whole course of the data with the client as part of relapse prevention, and pick out personal warning signs and effective responses [synthesis].
  3. End app access, as in ACT-DL [source: Myin-Germeys et al. 2021]. Alternatively, agree a self-managed continuation without clinician access.
  4. Settle the data: delete it, block your access, or keep the summaries in the clinical record [source for options: Mestdagh et al. 2023].
  5. Confirm that the client's crisis resources no longer depend on the app [synthesis].
• Example language (illustrative): "Over the next two weeks the prompts will thin out. Then we'll look at the whole picture together and decide what you want to keep, and whether your data should be deleted."
• Watch for: Loss of a felt connection when prompts stop. Data left on a platform without a decision. Clients who want to keep self-monitoring obsessively [synthesis].

5. Handling problems

The sources say little about clinical troubleshooting. Most of the guidance below is [synthesis] unless cited.

• Low or falling compliance.
  • Start by reading the pattern: which times of day are missed, and from which week.
  • Cut items first. Longer questionnaires raised burden and lowered data quality, while higher prompt frequency did not in students [source: Eisele et al. 2022].
  • Move prompts away from busy windows.
  • Reconnect the data to the client's own question.
  • Expect drop-off. In INTERACT, loss to follow-up after the intervention was 30% in the ACT-DL arm [source: Myin-Germeys et al. 2021].
  • Illustrative language: "The missing evening prompts tell us something too. What's going on at that time?"
• Burden or irritation.
  • Keep push content to the least needed, and drop components that don't shift anything [source: Walton et al. 2020].
  • Offer shorter blocks, as in ACT-DL's 3-day blocks [source: Myin-Germeys et al. 2021].
  • Offer pull-only use, or pause altogether.
• Careless or rote answering.
  • Suspect it when answers are identical across prompts or responses are very fast. Long questionnaires increase careless responding [source: Eisele et al. 2022].
  • Raise it without blame, then shorten or vary the items.
• Reactivity: monitoring makes things worse.
  • Rumination, checking or heightened symptom focus can follow repeated self-rating, especially in health anxiety, OCD or eating disorders [synthesis].
  • Reduce symptom items and add items on values or activities.
  • Use event-contingent sampling rather than many random prompts, or stop.
  • For eating disorders, avoid items on weight or food quantity unless that monitoring is part of the protocol and the eating-disorder team is involved [synthesis].
• Concealment or under-reporting of risk.
  • Knowing that a high answer triggers a call may lead some clients to under-report or stop responding. The review raises this as an untested concern [source: Bentley et al. 2021].
  • Discuss it openly, and explain why outreach exists.
  • Ask directly about risk in sessions whatever the app shows.
• Alarming entries outside your hours.
  • Follow the pre-agreed protocol (4.12) [source: Bentley et al. 2021].
  • If you discover an entry late, contact the client promptly, assess, document, and review whether your consent statement and review rhythm were accurate.
• Paranoia, suspicion or privacy fears.
  • Explain storage and access in concrete terms [source for platform facts: Mestdagh et al. 2023].
  • Offer pseudonymous accounts, no passive sensing, or paper exercises [source for paper option: Balaskas et al. 2021].
  • In psychosis, frame the app as the client's own tool and let them control what is shared [synthesis].
• Coercive or shared-device situations. Do not use EMI where a partner or family member could read entries and put the client at risk. Choose paper or session-only work [synthesis].
• Data contradicting the client's account (rupture risk).
  • Present the discrepancy with curiosity, let the client interpret it, and accept that either source may be wrong.
  • Repair if the client feels caught out: "I didn't mean to use the graph against you."
  • Momentary and recalled reports often differ for ordinary reasons [source: Stone et al. 1998].
• Stuck client: data collected but no insight.
  • Return to a narrower hypothesis, and add one context or behaviour item [source for hypothesis-driven design: von Klipstein et al. 2023].
  • Alternatively, switch from EMA to prompted skill practice.
• Non-adherence to exercises (answers prompts, skips the exercise).
  • Check whether the skill was learned in session, and whether the exercise is short enough to do in the moment.
  • Let the client choose the format: audio, text or image [source for formats: Balaskas et al. 2021].
• Technical failures.
  • Check whether battery optimisation is killing the app, whether the device is supported, and whether the client is online [source: Mestdagh et al. 2023].
  • Agree a paper back-up, and use the recovery code to restore data after a phone change [source: Mestdagh et al. 2023].
• Over-reliance on the app.
  • Build in tapering early, and favour pull use as therapy progresses.
  • Discuss what the client will do without prompts.
• Clinician overload.
  • Limit the number of clients on active EMI at once, and use templates for item sets.
  • Never include risk items you cannot review.
• Pressure to act on algorithmic signals. Decline to make clinical decisions from unvalidated sensor or machine-learning outputs (4.13) [source for development status: Balaskas et al. 2021].
• Medication questions arising from the data. For example, "my mood drops after my dose". Encourage the client to raise it with the prescriber and, with consent, pass the observation on. Never advise a change [synthesis].
• Minors and families.
  • Agree the confidentiality and parental-contact rules before starting [source: Bentley et al. 2021].
  • Check the platform's age terms and local consent rules [uncertain: not verified for specific platforms].

Sources

Primary and secondary sources:

• Balaskas A, Schueller SM, Cox AL, Doherty G (2021). Ecological momentary interventions for mental health: A scoping review. PLoS ONE 16(3): e0248152 (UCL Discovery PDF)
• Myin-Germeys I, van Aubel E, Vaessen T, et al. (2021). Efficacy of Acceptance and Commitment Therapy in Daily Life (ACT-DL) in early psychosis: results from the multi-center INTERACT randomized controlled trial. medRxiv preprint, doi:10.1101/2021.05.28.21257986
• Mestdagh M, Verdonck S, Piot M, et al. (2023). m-Path: an easy-to-use and highly tailorable platform for ecological momentary assessment and intervention in behavioral research and clinical practice. Frontiers in Digital Health 5:1182175
• von Klipstein L, Servaas MN, Schoevers RA, van der Veen DC, Riese H (2023). Integrating personalized experience sampling in psychotherapy: A case illustration of the Therap-i module. Heliyon 9(3): e14507 (DOAJ record; abstract)
• Bentley KH, Maimone JS, Kilbury EN, et al. (2021). Practices for monitoring and responding to incoming data on self-injurious thoughts and behaviors in intensive longitudinal studies: A systematic review. Clinical Psychology Review 90: 102098 (Nock Lab copy)
• Eisele G, Vachon H, Lafit G, et al. (2022). The effects of sampling frequency and questionnaire length on perceived burden, compliance, and careless responding in experience sampling data in a student population. Assessment 29(2): 136–151 (Maastricht University record; abstract)
• Walton AE, Collins LM, Klasnja P, Nahum-Shani I, et al. (2020). The micro-randomized trial for developing digital interventions: experimental design considerations. arXiv:2005.05880 (draft, not peer reviewed)
• Nahum-Shani I, Smith SN, Spring BJ, et al. (2018). Just-in-time adaptive interventions (JITAIs) in mobile health: key components and design principles for ongoing health behavior support. Annals of Behavioral Medicine 52(6): 446–462 (Penn State record; abstract)
• Myin-Germeys I, Klippel A, Steinhart H, Reininghaus U (2016). Ecological momentary interventions in psychiatry. Current Opinion in Psychiatry 29(4): 258–263 (Open Universiteit record; bibliographic record only)
• Stone AA, Schwartz JE, Neale JM, Shiffman S, et al. (1998). A comparison of coping assessed by ecological momentary assessment and retrospective recall. Journal of Personality and Social Psychology 74(6): 1670–1680 (Duke Scholars record; abstract)

Cited via secondary source:

• Heron KE, Smyth JM. Ecological momentary interventions: incorporating mobile technology into psychosocial and health behaviour treatments. British Journal of Health Psychology (online 2009), doi:10.1348/135910709X466063. The definition of ecological momentary intervention is used here as summarised by Balaskas et al. (2021).

Trial protocol (abstract):

• Riese H, von Klipstein L, Schoevers RA, van der Veen DC, Servaas MN, et al. (2021). Personalized ESM monitoring and feedback to support psychological treatment for depression: a pragmatic randomized controlled trial (Therap-i). BMC Psychiatry, doi:10.1186/s12888-021-03123-3 (Epistemonikos record)