CalmReminder: Watch-Based Probe for ADHD Parenting

• CalmReminder is a watch-based design probe that detects child calmness via personalized IMU data analysis to support timely positive reinforcement.
• The system employs a simple linear regression on motion energy, achieving a 78% calm detection rate when notifications are delivered during calm moments.
• Field deployment with 16 families revealed that parents appropriated notifications for mindfulness, planning, and conversations, adapting the system to their unique parenting styles.

CalmReminder is a parent-facing, watch-based design probe for families raising children with ADHD/hyperactivity. It uses a child’s smartwatch to detect moments of relative calm from motion data and then delivers just-in-time prompts to parents, with the aim of helping them notice, reflect on, and potentially reinforce calm behavior in the moment rather than responding only to disruption. In a four-week deployment with 16 families, of whom 12 completed the study, the system compared notification strategies ranging from hourly to random to only when the child was inferred to be calm; the reported findings emphasize both the feasibility of real-time calm detection and the fact that parents appropriated the notifications in multiple ways, including praise, mindfulness, activity planning, and conversation (Arakawa et al., 18 Feb 2026).

1. Definition, scope, and behavioral rationale

CalmReminder was introduced against a specific family dynamic: children with ADHD/hyperactivity often display frequent movement and impulsive behavior that disrupt daily routines, while parents may become stressed and reactive. The paper frames this as a feedback loop in which child dysregulation increases parent stress, and stressed parenting can worsen child behavior. Within that framing, a central challenge is attentional asymmetry: calm, desirable behavior is easy to miss, whereas disruptive behavior is salient and immediately action-forcing (Arakawa et al., 18 Feb 2026).

The intervention logic therefore centers on timely positive reinforcement. The paper states that parent training research suggests that timely positive reinforcement of calm or compliant behavior is more effective than only reacting to misbehavior. CalmReminder operationalizes that idea by trying to surface calm moments when they are happening, not retrospectively. Its design goals were explicit: to detect calm moments in real time, support parents in the moment, encourage positive reinforcement, avoid surveillance, and respect family variation. The system does not present raw motion data to parents; the watch is framed as a supportive sensing tool rather than a monitoring device. Just as importantly, the authors do not assume one correct usage pattern. CalmReminder is explicitly described as a design probe, meaning that families are expected to appropriate it in ways that fit their own parenting philosophies and routines (Arakawa et al., 18 Feb 2026).

That design-probe framing distinguishes CalmReminder from a fixed behavioral intervention. The notifications are not treated as prescriptions with a single intended action. Instead, the paper positions them as resources for parental noticing and reflection. This is significant because it relocates value from algorithmic correctness alone to the interaction between sensed calmness, parental interpretation, and family-specific practice.

2. Wearable sensing and calm detection pipeline

CalmReminder uses an Apple Watch to continuously collect IMU data, especially accelerometer data, and derives a motion feature described as motion energy over a 5-minute window. The feature is characterized as the root mean square of acceleration, or the energy of the acceleration signal. The compact form given in the study is

$$E_t = \sqrt{\frac{1}{N}\sum_{i=1}^{N} \lVert \mathbf{a}_i \rVert^2}$$

where $\mathbf{a}_i$ is the acceleration vector at sample $i$, $N$ is the number of samples in the 5-minute window, and $E_t$ is the energy estimate for that window (Arakawa et al., 18 Feb 2026).

The personalization layer is deliberately simple. For each participant, the system trains a linear regression model whose input is the average motion energy during the one-hour period before a parent survey response and whose output is the parent-reported perceived activity level on a 5-point scale, with $1 = \text{Very Calm}$ and $5 = \text{Very Active}$. The paper gives the model as

$\hat{y} = \beta_0 + \beta_1 E$

and defines the calm decision rule as

$\hat{y} < 3$

so a predicted value below 3 triggers a calm-only notification (Arakawa et al., 18 Feb 2026).

Several implementation details are central to the system’s practical form. The watch app runs on Apple Watch Series 7 or later. Motion features are computed locally and uploaded every 5 minutes rather than transmitting raw streams continuously. Battery life was about 8–10 hours per charge. A 4-digit PIN was required to stop recording. On the parent side, a phone app delivered notifications and surveys. Notifications expired after 30 minutes, end-of-day surveys after 12 hours, and end-of-week surveys after 2 days (Arakawa et al., 18 Feb 2026).

The architecture is technically lightweight, but its methodological role is substantial. By reducing the sensing pipeline to a personalized mapping from motion energy to parent-perceived activity, the system aligns its operational definition of calm with family-specific judgments rather than a generic motor threshold. This suggests a deliberate preference for participant-relative inference over an externally imposed calmness taxonomy.

3. Deployment design and notification conditions

The evaluation used a four-week within-subjects home deployment. Sixteen parent-child dyads were recruited, and 12 completed the study. The children were ages 6–12, with mean age 8.5, and 12 were male. All parents were mothers. Compensation was \$75. Most children were highly hyperactive on the Vanderbilt-related screening items; 12 of 16 scored above 10 out of 15 on the hyperactivity subset (Arakawa et al., 18 Feb 2026).

The study was organized into four weeks with distinct notification regimes. Week 1 was None, with no intraday notifications and only an end-of-day reflection survey; this served as both baseline behavior and acclimation. Week 2 was Hourly, with hourly reflection prompts from 8 AM to 8 PM; this week was also used to collect training data for personalization. Weeks 3 and 4 compared Random and Calm-only conditions in counterbalanced order. In both of those weeks, the system delivered at most five notifications per day and at most one per hour. The random and calm-only conditions were explicitly paired with positive reinforcement prompts asking parents to support or praise the child (Arakawa et al., 18 Feb 2026).

The evaluation combined quantitative and qualitative methods. The pre-study survey collected background information and baseline hyperactivity screening using 5 items from the NICHQ Vanderbilt scale. Intraday surveys, administered during Weeks 2–4, asked about child activity level on a 5-point scale and included open-text descriptions of recent activities. End-of-day surveys, administered in all weeks, covered medication adherence, parent-child communication effectiveness, and free-text reflection. End-of-week surveys included child behavior description, parental self-efficacy or overwhelm, parent-child relationship quality, technology evaluation, and open-text reflections on change and learning. Exit interviews addressed comparative evaluation of random versus calm-only notifications, daily-life impact, communication, feedback, and perceived changes in family interactions (Arakawa et al., 18 Feb 2026).

This protocol is notable for separating model training, baseline observation, and comparative prompting conditions, while also embedding the system in ordinary home routines rather than clinic-based tasks. The design therefore evaluates not only detection quality but also how notification strategies are lived with and interpreted over time.

4. Quantitative findings and the timing of prompts

The calm detection model produced a global model $R^2 = 0.25$ and a personalized model $\mathbf{a}_i$0. The paper characterizes this as moderate prediction, but operationally sufficient for the design probe. The key alignment result is that, when the system triggered calm-only notifications, 78.0% of responded notifications were reported by parents as actually calm. The comparison conditions were substantially lower: the Random condition had a calm ratio of 38.7%, and the Hourly condition had 42.9% in the data (Arakawa et al., 18 Feb 2026).

Response rates give additional evidence about burden and feasibility. Among completed families, intraday response rates were 59.6% ± 16.6 for Hourly, 59.0% ± 17.7 for Random, and 64.8% ± 20.0 for Calm-only. End-of-day response rates were 78.4% ± 24.5 for None, 74.1% ± 28.6 for Hourly, 62.5% ± 19.3 for Random, and 74.3% ± 18.1 for Calm-only (Arakawa et al., 18 Feb 2026).

Attrition was itself informative. Four families dropped out, mostly during the hourly week. The dropouts were clustered among parents with high baseline stress: 3 of the 4 dropouts had stress = 4, and the fourth had stress = 3. The paper interprets this as suggesting that high stress may prevent initial engagement, even if it does not determine later appropriation style among adopters (Arakawa et al., 18 Feb 2026).

The paper also complicates any simple interpretation of “good timing.” Parents often reported that calm-only notifications felt appropriately timed because they arrived when the child was relatively calm. But the study argues that appropriate timing was not reducible to objective calm detection. Parents also judged notifications as well-timed when they provided a moment to praise, a reminder to pause and reflect, a cue to be mindful of their own behavior, a prompt to plan activities, or a way to start a conversation. Conversely, some parents did not experience random or hourly prompts as poorly timed when those prompts were useful for reflection. A common misconception would therefore be that the value of CalmReminder is exhausted by classification performance. The deployment instead indicates that the meaning of timing is relational and situated, not purely algorithmic.

5. Appropriation, parental agency, and the design-probe interpretation

A central contribution of CalmReminder is the claim that parents were not passive recipients of notifications. The paper identifies several appropriation patterns. Some parents used the system in the intended manner, employing notifications for positive reinforcement during calm moments. Many used the prompts as a mindfulness or self-reflection cue, turning a child-centered alert into a reminder to ask themselves whether they had been patient, whether they were giving enough compliments, or whether they should speak more calmly. Others used the calm or activity signal for activity planning and hyperactivity management, treating the notifications as cues for scheduling, physical outlets, or routine management. At least one parent used the system as a conversation starter to explain ADHD and discuss why certain behaviors happen. There was also rejection: two parents felt the system added little value because they already had strong positive reinforcement routines (Arakawa et al., 18 Feb 2026).

The authors summarize these patterns as two broad modes. In amplifying existing strategies, CalmReminder intensified practices already present in the family, such as praise, intentional communication, or awareness of activity outlets. In expanding repertoires, it supported new practices, including mindfulness, new praise habits, or conversation routines that had not previously been part of family interaction (Arakawa et al., 18 Feb 2026).

This leads to the paper’s strongest conceptual claim: families actively re-designed CalmReminder. The system’s effect did not derive solely from a prompt being sent during a calm moment. Rather, value emerged because parents selected which notifications mattered, interpreted them through their own values, used them for goals not specified by the designers, and combined them with pre-existing parenting strategies. The implication is not that adaptation is a secondary implementation detail; it is constitutive of how the intervention works.

The design implications follow directly from that interpretation. The paper argues that parent-facing intervention systems should support flexible configuration, design for parent agency, support both child-focused and parent-focused value, treat family context as central, avoid overloading stressed families, support dyadic sensemaking, and regard wearables as relational artifacts rather than mere sensors (Arakawa et al., 18 Feb 2026).

6. Position within the broader reminder-systems literature

Although CalmReminder names a specific watch-based ADHD parenting system, adjacent work suggests a broader reminder design orientation organized around subtle prompting, contextual interpretation, and non-coercive support. In smartphone overuse research, for example, a notification-based intervention was designed to be “subtle and familiar,” to avoid interrupting the current workflow, to avoid being annoying or burdening, and not to forcefully restrict smartphone activity; in a 3-week experiment, average weekly smartphone usage fell from 41 hours 26 minutes 47 seconds in baseline to 38 hours 57 minutes 22 seconds in intervention, a 6.01% reduction, with 75% of participants reducing usage during the notification week (Sarker et al., 19 Jul 2025). This suggests a reminder-not-restriction logic closely aligned with CalmReminder’s emphasis on lightweight prompts rather than punitive control.

In dementia care, Remindful extends task prompting with caregiver-facing alerts, summaries, and review features, but explicitly argues that reminder systems should not be treated as neutral behavioral sensors. Reminder interaction data were found to be highly context-dependent, shaped by household participation, prompt attribution, routine mismatch, accessibility barriers, and technical failures; the system is therefore framed as an assistive infrastructure for caregiver interpretation that preserves uncertainty (Lai et al., 21 Apr 2026). That position resonates with CalmReminder’s finding that notification value depends on family interpretation rather than on a one-to-one mapping from sensed state to behavioral meaning.

Work on intentional digital living provides a complementary reminder model in desktop computing. The Intent Assistant elicits a user’s intention, clarifies it through up to two follow-up questions, monitors screenshots, application titles, and URLs every 2 seconds, and delivers gentle nudges when behavior diverges from the stated goal. In a three-week, within-subjects field study with 22 participants, it was compared against a simple periodic reminder system and a logging-only baseline, and its technical evaluation on IntentionBench reported accuracy: 0.878, precision: 0.959, recall: 0.755, and F1: 0.845 (Choi et al., 16 Oct 2025). Relative to CalmReminder, this line of work shows how “gentle” reminder systems can be made context-aware without becoming hard blockers, though it also foregrounds notification burden and privacy concerns.

Smart-home reminder research pushes the same agenda toward richer contextual authoring. A conversational pipeline for authoring reminders in everyday language compiles user requests into Boolean Python functions over current time, sensor state, activity data, and a shared memory structure called a blackboard. Across two studies, overall correct code generation increased from 45.5% in Study 1 to 76.7% in Study 2, supporting time-based, activity-based, sensor-based, and state-machine triggers (Chan et al., 21 May 2026). A plausible implication is that CalmReminder’s calm-triggered parent prompts occupy one point in a wider space of context-aware reminders whose execution increasingly depends on sensed state rather than clock time alone.

Finally, evaluation work on LLMs frames reminder-like behavior as a prospective-memory problem rather than a retrospective-memory problem. TriggerBench defines prospective memory as the ability to notice a later trigger, recall an earlier latent constraint, and intervene proactively without being explicitly asked. Across 1,265 PM tasks and 440 RM probes, the benchmark shows that proactive reminder behavior is harder than direct retrieval, exhibits a precision-recall trade-off, and is fragile under overloaded triggers (Zhang et al., 22 Jun 2026). This broader result does not concern CalmReminder directly, but it sharpens one of its underlying design stakes: reminder systems are not only about detecting states; they are about deciding when intervention is warranted and when silence is preferable.

Taken together, these adjacent literatures suggest that CalmReminder is best understood not merely as a wearable classifier for child calmness, but as a concrete instantiation of a larger research movement toward calm, context-sensitive, autonomy-preserving reminder systems. Within that movement, its distinctive contribution is to show that the parent, not the notification policy alone, is the decisive locus of interpretation and action.