PaperlayerView exact revision on arXiv

Source-linked AI summary

Guided Reflection for Personal Sleep Insight in Everyday Sleep Tracking

Bokyung Kim, Amama Mahmood, Honghao Zhao, Molly E. Atwood, Luis F. Buenaver, Ziang Xiao, Chien-Ming Huang

arXiv:2610.10822v1cs.HC

TL;DR

Sleep trackers make changes visible, but users may have difficulty interpreting what those changes mean in their own lives. DREAM adapts guided discovery into a voice-based system that uses users’ interpretations to shape education. In a six-week comparison with generic education, DREAM participants reported greater personal sleep insight, motivation, and willingness to use the system.

  • Problem

    Sleep tracking can reveal changes without resolving how users should interpret them in the context of their own lives, while reflection and guidance in digital systems are often treated separately.

    Source: §1 · p. 2

  • Method

    DREAM is an LLM-powered voice system that selectively discusses sleep changes, elicits users’ interpretations, and uses them to determine whether and what education should follow.

    Source: §1 · p. 3

  • Results

    Compared with Generic Education, DREAM participants reported greater personal sleep insight, motivation, and willingness to use the system, and more often reported trying a sleep-related strategy.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Guided reflection helped participants connect sleep changes to their circumstances and gave them a clearer rationale for strategies they tried.

    Source: §9 · p. 24

    The field study did not isolate guided reflection from personalization, because DREAM combined reflection with tailored education while Generic Education was independent of sleep data.

    Source: §8.5 · p. 24

Abstract

from arXiv · show

Digital sleep technologies make tracking accessible, yet users often struggle to interpret what changes in their sleep mean. Behavioral sleep medicine addresses this through guided discovery, helping patients develop personal interpretations rather than simply receiving explanations. To bring this to everyday tracking, we present DREAM, an LLM-powered voice assistant that monitors conversational sleep diaries, selectively invites users to interpret meaningful changes, and uses their interpretation to tailor subsequent education. We co-designed DREAM with sleep specialists iteratively and evaluated it in a six-week field study (N=14) against a generic-education control. DREAM participants reported greater personal sleep insight, motivation, and willingness to use the system, and described a clearer rationale for trying strategies. Our findings suggest that guided reflection made participants more active interpreters of their own sleep. Furthermore, we argue that expert involvement is not a single transfer of knowledge but an ongoing process of making tacit judgment explicit and testable.

1 Introduction

DREAM addresses the gap between noticing sleep changes and understanding them in the context of everyday life. It uses guided reflection to connect users’ interpretations to subsequent education, and the field study reported stronger insight and motivation than generic education.

1 Introduction

  • Digital tracking can make sleep changes visible without clarifying their meaning or how to respond in the context of one’s own life.The interpretive challenge is to bridge observation and knowing what to do next.Source: §1 · p. 2
  • The authors adapt guided discovery and sleep-diary review as design resources for everyday users, without attempting to reproduce clinical care.They frame this as support for people seeking to understand sleep in everyday settings where access to behavioral sleep medicine is limited.Source: §1 · p. 2
  • DREAM selectively brings changes in a user’s sleep record into conversation, asks what may have contributed, and uses the user’s interpretation to determine whether and what education should follow.The paper describes the system as an LLM-powered, voice-based assistant integrating conversational sleep diaries, longitudinal monitoring, guided reflection, and sleep education.Source: §1 · p. 3
  • Compared with Generic Education, DREAM participants reported greater personal sleep insight, motivation, and willingness to use the system, and more often reported trying a sleep-related strategy.Interviews described how reflection connected sleep changes to personal circumstances and made subsequent guidance meaningful.Source: §1 · p. 3
  • The paper presents design implications for using LLMs to elicit interpretations of personal data and connect them to tailored subsequent support.The contributions include an empirical study of guided reflection and implications for LLM-powered behavioral health systems.Source: §1 · p. 3

2 Related Work

Related work spans sleep monitoring, reflection on personal data, conversational support, and adaptive digital sleep systems. Across these areas, the paper identifies a gap in supporting users to interpret specific changes and connecting that interpretation to subsequent guidance.

2.1 Sleep Monitoring and the Interpretive Gap

  • Sleep tracking makes longitudinal records more accessible, but users may still struggle to interpret metrics, judge their significance, and identify changes to make.More detailed records do not inherently clarify what contributed to a change or what to do in response.Source: §2.1 · pp. 3–4

2.2 Supporting Reflection on Personal Health Data

  • Personal informatics frames reflection as relating data to lived experience to develop self-knowledge, yet tracking does not consistently lead to insight or behavior change.How systems can scaffold interpretation from a specific observation remains underexplored.Source: §2.2.1 · p. 4, §2.2 · p. 4

2.2.1 Reflection in Personal Informatics and Behavioral Sleep Medicine.

  • Behavioral sleep medicine uses diary review and collaborative questioning to examine sleep patterns and possible contributors, adapting strategies as understanding develops.DREAM adapts this questioning stance to help users interpret sleep-record changes in relation to their experiences, rather than restructure sleep-related beliefs.Source: §2.2.1 · p. 4

2.2.2 Conversational and LLM-Based Support for Reflection.

  • Prior conversational and LLM systems personalize reflection or guidance, but less often help users understand a change in personal data before guidance is provided.Examples include tracker-data narratives, adaptive journaling questions, cognitive reframing, and health coaching drawing on user context.Source: §2.2.2 · p. 4

2.3 Personalized and Adaptive Support in Digital Sleep Systems

  • Digital sleep systems shape what support is provided and when through personalized and adaptive approaches.The related work considers both reflection and guidance in these systems.Source: §2.3 · p. 4, §2.3.1 · p. 5

2.3.1 Connecting Reflection and Personalized Guidance.

  • Sleep reflection tools and personalized-guidance systems often operate separately, with guidance based on tracked data and context rather than the user’s interpretation.DREAM connects them by using the interpretation of a selected sleep change to determine whether guidance is warranted and what it should address.Source: §2.3.1 · p. 5

2.3.2 Deciding When to Support Reflection.

  • Selective reflection draws on longitudinal sleep data to identify observations worth discussing, rather than prompting reflection on every recorded night.The rationale includes avoiding prompts that can become repetitive or burdensome in repeated daily tracking.Source: §2.3.2 · p. 5

Research Overview

DREAM adapts guided discovery from behavioral sleep medicine for everyday sleep tracking, while explicitly not aiming to reproduce clinician-led CBT-I or provide diagnosis or treatment. The system was developed iteratively with specialists and refined before a six-week field study.

Research Overview

  • DREAM uses longitudinal sleep data to support interpretation of sleep changes and provides education when relevant, drawing on guided discovery and sleep-diary review as design resources.The authors distinguish this aim from reproducing clinician-led CBT-I, diagnosis, treatment, or automated clinical interpretation.Source: Research Overview · p. 5
  • Two behavioral sleep medicine specialists helped develop the system through specialist-informed design, implementation, a formative probe, refinement, and field study.The specialists had 6 and 21 years of clinical experience.Source: Research Overview · p. 5
  • The five-stage process used a 10-day technology probe to identify refinements before deploying the refined system in a six-week field study with 14 participants.The probe incorporated participant experiences, researcher checks, and specialist review.Source: Figure 2 · p. 6

3 Stage 1: Envisioning the System with Sleep Specialists

Sleep specialists and researchers translated behavioral sleep medicine practices into a workflow for monitoring, reflection, and education. Specialist walkthroughs of simulated cases helped refine how changes are selected, how interpretation is supported, and when guidance follows.

3 Stage 1: Envisioning the System with Sleep Specialists

  • The design process began with three objectives: selectively review longitudinal sleep patterns, guide users’ own interpretations, and ground behavioral guidance in that understanding.Two behavioral sleep medicine specialists helped translate these objectives into an everyday interaction.Source: §3 · p. 6, §3.1 · p. 6

3.1 Design Objectives

  • Reflection was designed to focus on selected changes, help users interpret them through their experiences, and base subsequent guidance on the resulting understanding.These objectives define monitoring, reflection, and education as the workflow’s three phases.Source: §3.1 · p. 6

3.2 Procedure

  • Specialists reviewed monitoring criteria, diary content, a reflection decision tree, and when and how education should follow; simulated cases further refined the logic before implementation.Walkthroughs considered whether a change merited reflection, how the conversation should proceed, and what guidance should follow.Source: §3.2 · p. 7

3.3 Findings

  • The specialist-informed findings specified decisions across three workflow phases: monitoring, reflection, and education.Monitoring pairs sleep outcomes with context, uses an individual’s recent sleep pattern as a baseline, considers both personal deviation and issue magnitude, and prioritizes one change for reflection.Source: §3.3 · p. 7, §3.3.1 · p. 7

3.3.1 Phase 1: Monitoring .

  • Monitoring pairs sleep measures with daily context and judges changes against an individual’s recent pattern, distinguishing personal deviations from absolute threshold crossings.When multiple measures change, the priority order is SOL/WASO, then TST, TIB, SE, NWAK, and sleep quality.Source: §3.3.1 · p. 7

3.3.2 Phase 2: Reflection .

  • Reflection begins by eliciting the user’s own account, then offers progressively more support when they are uncertain.Prompts progress from recalling differences in the prior day or routine, to contextual cues, then whether the change is isolated or recurring.Source: §3.3.2 · p. 7

3.3.3 Phase 3: Education .

  • Education is selected based on factors surfaced during reflection, connecting a relevant factor to the sleep issue and offering one or two concrete actions.Sleep-hygiene material supplied by specialists grounds the educational content, while the user’s interpretation determines what is relevant.Source: §3.3.3 · p. 8

4 Stage 2: Initial Prototype

The initial DREAM prototype converted conversational diaries into longitudinal records, monitored sleep changes against personal baselines, and used selected changes to guide reflection and education.

4 Stage 2: Initial Prototype

  • DREAM used backend rules for sleep calculations, baseline construction, change detection, and selection, while GPT-5.2 supported diary intake, extraction, reflection, and response generation.The prototype used separate monitoring, reflection, and education modules.Source: §4 · p. 8
  • Morning and evening diary conversations became structured longitudinal records containing sleep measures and daily context.The backend derived TIB, TST, and SE and maintained moving baselines for seven sleep measures.Source: §4 · p. 8
  • Monitoring evaluated each morning record against specialist-informed thresholds that combined personal-baseline deviations with absolute sleep criteria.Detected changes were flagged and, when multiple measures were flagged, one was selected for reflection.Source: §4 · p. 8
  • Reflection began by asking for the user’s interpretation, then used recall prompts, contextual cues, and frequency assessment if uncertainty remained.The reflection flow used the selected change, recent sleep pattern, and relevant evening context.Source: §4 · pp. 8–9
  • In the initial prototype, every selected flag initiated reflection, and every completed reflection automatically led to education.Education provided a brief explanation and one or two strategies from specialist-provided sleep-hygiene material.Source: §4 · p. 9

5 Stage 3: Technology Probe and Formative Review

The 10-day technology probe tested repeated use and surfaced shortcomings in when and how the prototype initiated reflection and guidance. Participant, researcher, and specialist feedback informed four refinements.

5.1 Participants and Procedure

  • During the 10-day home probe, 11 students completed morning and evening diaries; Days 1–3 initialized baselines, and Days 4–10 could lead to reflection and education.Participants completed post-study questionnaires and interviews.Source: §5 · p. 9, §5.1 · p. 10

5.2.1 Participant Interviews.

  • Some participants found extended reflection unnecessary when a change had an obvious explanation, and repetitive when similar routines produced similar interactions.These accounts suggested that detecting a change alone was not enough reason to continue reflection.Source: §5.2.1 · p. 10

5.2.2 Researcher Checks.

  • At the interaction level, 47 of 49 interactions (95.9%) passed all researcher checks; two contained extraction or interpretation errors that produced an incorrect sleep value and flag.Five interactions received independent review by both researchers, with agreement on all checks; the small overlap limits how to read that agreement.Source: §5.2.2 · p. 10

5.2.3 Specialist Review.

  • Specialists judged 22 of 49 flagged interactions (44.9%) appropriate and 27 (55.1%) inappropriate, despite high Noticing and Reasoning ratings.Action ratings varied more; the results indicated that the prototype did not yet reliably determine when reflection was warranted.Source: §5.2.3 · p. 11
  • Specialists identified ordinary variability, premature causal conclusions, and automatic educational tips as reasons the initial logic could misfire.They cautioned against over-intervening instead of connecting guidance to what was actually happening.Source: §5.2.3 · p. 11

5.3 Design Refinements

  • The refined monitoring separated detecting a change from deciding whether it warranted reflection, and stabilized longitudinal references against unusual values and outlier nights.Measures were screened independently before entering the reference data.Source: §5.3 · p. 11
  • Reflection was revised to clarify ambiguous or inconsistent explanations before drawing connections; education became conditional on what reflection established.When further support was unnecessary, the interaction could end with a brief acknowledgment or without guidance.Source: §5.3 · p. 11

6 Stage 4: Refining DREAM

DREAM’s refined workflow separated detected changes from reflection eligibility and made reflection determine whether education followed. The refinements also stabilized baselines and clarified how the system handled uncertainty and one-time explanations.

6 Stage 4: Refining DREAM

  • The refined pipeline retained conversational diaries and three modules but added an eligibility check before reflection, which could lead to education or a brief acknowledgment and exit.The figure distinguishes backend-rule steps from LLM-powered conversation.Source: §6 · p. 11, Figure 4 · p. 12

6.1.1 Stable Baseline Updating.

  • Week 1 established the baseline pool; later observations were evaluated against the cumulative record before that week and added only after it ended.Individual morning measures were screened independently, excluding a value when its deviation from the baseline median exceeded max(2 × MAD, metric-specific floor); evening context measures used 2 × MAD.Source: §6.1.1 · p. 12

6.1.2 Candidate Detection and Reflection Eligibility.

  • Candidate detection retained soft personalized, hard personalized, and literature-informed thresholds, while a separate backend gate determined reflection eligibility.A candidate was a detected change before the eligibility decision, not an automatic trigger for reflection.Source: §6.1.2 · pp. 12–13
  • Some related-measure patterns were immediately eligible, while weaker standalone changes could be suppressed or require recurrence within seven days.The patterns included fragmentation, efficiency, and sleep opportunity; a remaining soft candidate became eligible after a second occurrence of the same measure within seven days.Source: §6.1.2 · p. 13
  • When multiple candidates were eligible, related-measure patterns took priority; otherwise, stronger candidates preceded softer ones, followed by the prototype’s measure ordering.Candidates not yet eligible remained in the longitudinal record for recurrence checks, while the interaction ended after the diary.Source: §6.1.2 · p. 13

6.2 Refinements to the Reflection Module

  • Reflection presented a change neutrally and asked for the user’s view; ambiguous or inconsistent explanations prompted clarification before contextual cues were offered as possible, not inferred, causes.A clear one-time circumstance could instead receive acknowledgment and end the interaction without further probing.Source: §6.2 · p. 13

6.3 Refinements to the Education Module

  • Education followed when reflection established an actionable contributor connected to the sleep issue; reflection could also end with acknowledgment or without guidance.Reflection therefore determined both what education addressed and whether it was warranted; the illustrated interactions show these different outcomes.Source: §6.3 · pp. 13–14, Table 2 · p. 14

7.1 Study Design and Conditions

In a six-week field study, DREAM was compared with generic education, with participants sharing voice diaries and sleep-tracking tools. DREAM participants reported greater personal sleep insight, motivation, and willingness to use the system; interviews suggested they more often connected sleep changes to their own circumstances and strategies.

7.1 Study Design and Conditions

  • The six-week, between-subjects study compared DREAM with generic education after a shared one-week baseline; 14 adults completed the study.Both conditions used the same diaries, reminders, web application, and sleep-hygiene material; DREAM reflection and education were selective, while control received a rotating tip after each morning diary.Source: §7.1 · p. 14, §7.2 · pp. 15–16

7.3 Measures and Analysis

  • Measures covered system experience, personal sleep understanding, knowledge, motivation and uptake, while broader sleep and behavioral outcomes were treated as exploratory.The analysis included between-condition tests, longitudinal models for repeated questionnaire measures, and descriptive summaries of broader outcomes.Source: §7.3 · pp. 16–17, §7.3.3 · p. 17, §7.3.4 · p. 18

7.4.1 System Evaluation.

  • 52 of 60 refined-system interactions (86.7%) were judged appropriate for reflection, versus 22 of 49 (44.9%) in the technology probe.Noticing and Reasoning were rated highly in both deployments, suggesting the main improvement was identifying when reflection should occur.Source: §7.4.1 · p. 18

7.4.2 System Use and Experience.

  • DREAM participants reported greater willingness to use the system (p = .015, d = 1.33) and disclosure comfort (p = .044, d = 1.00).Mean SUS scores were 69.29 (SD = 16.25) for DREAM and 63.21 (SD = 21.25) for control; DREAM morning interactions also lasted longer on average.Source: §7.4.2 · p. 18

7.4.3 Sleep Understanding and Action.

  • DREAM participants reported greater personal sleep insight (p = .027, r_rb = .61), insight depth (p = .022, r_rb = .57), motivation (p = .035, d = 1.06), and behavioral uptake (p = .036, r_rb = .53).Personal specificity, relational reasoning, strategy specificity, and outcome reflection did not differ significantly between conditions.Source: §7.4.3 · p. 20
  • Sleep-hygiene awareness and caffeine knowledge increased across both conditions, but general sleep-knowledge measures showed no significant condition-specific differences.Sleep Beliefs and SHAPS knowledge likewise showed no condition-specific differences.Source: §7.4.3 · p. 20

7.4.4 Sleep and Behavioral Outcomes.

  • Sleep quality, sleep-hygiene practices, nightly measures, and recorded behaviors showed no significant condition-specific differences over six weeks.Across both conditions, total sleep time and sleep efficiency increased over time, while wake after sleep onset decreased.Source: §7.4.4 · p. 20

7.4.5 Participant Perspectives.

  • Interviews suggested diaries focused attention on sleep in both conditions, while DREAM participants linked observed changes to personal circumstances, valued selective reflection, and described education as informing strategies.Participants questioned reflection when a change did not seem meaningful; DREAM participants more often described actions as following their own interpretation.Source: §7.4.5 · p. 21

8 Discussion

The discussion presents guided reflection as a way to make sleep data personally meaningful and connect users’ interpretations to actionable education. It also frames specialist input as an iterative process of making tacit judgments explicit, while noting limits in the sample, study comparison, and system safeguards.

8.1 Making Sleep Data Personally Meaningful through Guided Reflection

  • DREAM participants reported greater personal sleep insight and willingness to continue despite receiving education less frequently, as reflection invited their interpretations before guidance.Participants also reported higher comfort disclosing information; the authors suggest that seeing their responses shape the interaction may make disclosure feel more meaningful.Source: §8.1 · p. 22

8.2 Connecting Personal Understanding to Action

  • Guidance grounded in a participant-identified contributor gave them a clearer rationale for trying a strategy than generic tips disconnected from their own sleep.Participants described moving from recognizing patterns to knowing how to take action, consistent with guidance following their own interpretation.Source: §8.2 · pp. 22–23

8.3 Design Implications for LLM-Supported Reflection in Personal Health

  • DREAM uses an LLM to present an observation, ask what may have contributed, clarify the user’s account, and then decide whether guidance should follow.The authors suggest this process may extend to other personal-health tracking and behavioral-health contexts when domain experts define appropriate observations, questions, and guidance.Source: §8.3 · p. 23

8.4 Translating Specialist Judgment into Adaptive System Logic

  • After specialist judgments were incorporated into the refined system, 86.7% of reviewed interactions were judged appropriate for follow-up, up from 44.9%.The authors frame expert involvement as progressively making tacit judgments explicit and testable, rather than transferring expertise once into prompts or rules.Source: §8.4 · pp. 23–24

8.5 Limitations and Future Work

  • The small, primarily university-based sample limits demographic diversity and the generalizability of findings across daily routines and tracking needs.The authors call for larger, more diverse deployments.Source: §8.5 · p. 24
  • Alexa speech-recognition errors and command misinterpretations could disrupt longer reflection conversations and may have affected engagement.The authors suggest exploring more flexible conversational platforms, including mobile interfaces supporting voice and text.Source: §8.5 · p. 24
  • Because the field study combined guided reflection with issue-tailored education, it cannot isolate reflection’s effect from other aspects of personalization.The authors suggest comparing reflection with personalized education delivered without reflection, or varying the two components independently.Source: §8.5 · p. 24
  • Although education used specialist-provided sleep-hygiene material, DREAM lacked explicit mechanisms to detect hallucinated, unsupported, or clinically inappropriate responses.The authors call for stronger safeguards, clear scope boundaries, and pathways to appropriate clinical support.Source: §8.5 · p. 24

9 Conclusion

DREAM applies guided discovery to everyday sleep tracking by helping users interpret changes in their own sleep and grounding education in that understanding. In a six-week field study, participants showed deeper personal understanding and a clearer rationale for strategies they tried than participants receiving generic education.

9 Conclusion

  • DREAM combined selective discussion of sleep-record changes, support for users’ own interpretations, and education grounded in that understanding.In a six-week field study with 14 participants, DREAM was compared with a generic-education control.Source: §9 · p. 24

Declaration of generative AI and AI-assisted technologies in the writing process

The authors report using ChatGPT and Claude during preparation to reduce repetition and improve readability and language. They also used Claude to help design figures and generate figure descriptions, then reviewed and edited the content.

Declaration of generative AI and AI-assisted technologies in the writing process

  • The authors used ChatGPT and Claude to reduce repetition and improve language, and Claude to help design figures and generate figure descriptions.They state that they reviewed and edited the material and take responsibility for the publication’s content.Source: Declaration of generative AI and… · p. 25

A.1 Platform, Architecture, and State Management

The appendix describes DREAM’s deployment and conversational modules, along with its diary, coding, questionnaire, outcome, and specialist-evaluation procedures. It reports that questionnaire trajectories and recorded-behavior changes did not differ significantly between conditions.

A.1 Platform, Architecture, and State Management

  • DREAM ran as an Alexa voice application with AWS Lambda middleware, Azure-hosted backend logic, GPT-5.2 conversational components, and participant-specific Google Sheets records.A dual-API middleware architecture coordinated Alexa interactions with backend processing under Alexa’s response-time constraints.Source: §A.1 · p. 29

A.2.1 Diary Intake Conversation Handler / A.2.2 Text-to-JSON Converter

  • Separate prompts handled conversational diary intake and conversion of transcripts into structured JSON fields for backend metric derivation.The diary handler asked predefined questions conversationally and could clarify ambiguous responses.Source: §A.2.1 · p. 29, §A.2.2 · p. 29, §A.2 · p. 29

A.2.3 Reflection Handler / A.2.4 Education Handler

  • The reflection handler received flagged deviations and relevant evening context, while the education handler linked users’ interpretations to actionable sleep-hygiene strategies.Reflection prompts instructed the model to ask open-ended questions first and offer context as cues, not asserted causes; education prompts forbade diagnostic claims beyond the provided material.Source: §A.2.3 · p. 29, §A.2.4 · p. 29

B Sleep Diary Questions

  • Morning and evening diaries had fixed required content, with conversational phrasing and clarification questions available for ambiguous or incomplete responses.The LLM phrased each predefined diary question conversationally.Source: B Sleep Diary Questions · p. 29

C Open-Ended Response Coding Criteria

D Post-Study Questionnaire Items

E Sleep and Behavioral Outcomes

  • No questionnaire showed a significant condition-by-time interaction, and no between-condition difference in longitudinal change for recorded behaviors reached statistical significance.The appendix identifies separate tables for repeated questionnaire measures, nightly sleep outcomes, and exploratory behavior changes, without giving their cell values here.Source: E Sleep and Behavioral Outcomes · p. 31, Table 8 · p. 31, Table 9 · p. 31

F Detailed Specialist Evaluation

  • Specialist evaluation covered system flagging and stage-level ratings, with stage samples varying when specialist ratings were unavailable.Noticing and Reasoning were assessed for interactions judged appropriate for follow-up; Action was assessed only when interactions proceeded to education. Stage-level ratings used five-point scales.Source: F Detailed Specialist Evaluation · p. 32, Table 11 · p. 32
Loading 2610.10822v1…