"Can Computing Systems Eventually Disappear?": A Long-Term Tracking Study on the Use of Diabetes Computing Systems by Patients with Diabetes

1 Introduction

• The study follows 37 patients across five waves from 2017 to 2025, with 15 practitioners interviewed to examine user-system change and its design conditions.Its aim is process-level mechanism, not population-level generalization.Source: §1 · p. 2
• Continued use with cognitive disengagement describes users continuing to use systems while self-initiated interpretation recedes.The introduction distinguishes this state from abandonment and locates it at the end of a three-stage trajectory.Source: §1 · p. 2
• Clinical prioritization, compressed information, and automation reinforce systems that substitute for users’ understanding, shifting burden from understanding to responding.The reported consequences include less independent judgment of bodily state and fatigue rather than accumulated competence.Source: §1 · p. 2
• The paper proposes layered explanations, preserved user context, and failure-aware design to maintain interpretive engagement while automating routine work.These implications accompany a broader argument to shift ubiquitous health computing from judgment substitution toward preserved interpretation.Source: §1 · p. 2
• Whether similar disengagement occurs beyond diabetes remains an open empirical question, despite possible relevance to other ubiquitous systems.The authors position their findings as a precedent for future research, not a tested cross-domain result.Source: §1 · pp. 2–3

2 Related Work

• The paper links five research strands through the question of what happens to users’ interpretation, autonomy, and long-term health management as systems assume interpretive work.The strands are ubiquitous computing, human–automation interaction, wearable diabetes technologies, long-term self-tracking, and disengagement.Source: §2 · p. 3

2.1 Ubiquitous Computing, Cognitive Offloading, and Human–Automation Interaction

• Prior critiques of technological disappearance are conceptually developed but empirically thin on how consequences unfold as practices and capabilities co-evolve.The study addresses this with longitudinal evidence without assuming that disappearance necessarily produces disengagement.Source: §2.1 · pp. 3–4
• Longitudinal evidence remains limited on how sustained offloading changes interpretive practices as systems grow more capable.Cognitive disengagement means continued use becomes decoupled from self-initiated sensemaking; unlike offloading, it need not entail an error or inappropriate reliance.Source: §2.1 · p. 4

2.2 Long-Term Personal Informatics and Wearable Diabetes Technologies

• Research on diabetes technologies and personal informatics shows that automation redistributes work and tracking practices change, but does not trace interpretation across successive technology generations.Existing frameworks leave a conceptual gap: continued use can coexist with markedly less frequent interpretation, model-building, and questioning.Source: §2.2 · pp. 4–6

2.3 Potential Risks of Cognitive Disengagement in Wearable Health Systems

• Potential risks include judgment failures, fatigue, reduced interpretive readiness, and weaker recovery capacity, but these are plausible interacting consequences rather than inevitable outcomes.Existing work rarely studies how such risks connect over years as systems move from displaying data to interpreting, predicting, and acting.Source: §2.3 · pp. 6–7

3.1 Study Design

• The longitudinal design follows the same primary participants at five time points to trace changes in usage contexts, interpretation, and decision-making.One-time practitioner interviews supplement the user trajectories with perspectives on how systems are constructed and use is preconfigured.Source: §3.1 · p. 7

3.2 Participants

• The primary group comprises 37 people with varied ages, gender, rural or urban settings, education, and occupations; 15 practitioners add five doctors, five designers, and five vendors.Primary participants were recruited through in-situ contact and snowball sampling, while practitioners were purposively sampled with balanced roles.Source: §3.2 · p. 7

3.3 Ethics

• Participants gave informed consent, could withdraw, and interviews used non-leading questions; data were anonymized, securely stored, and used for academic research.The study received institutional ethics approval.Source: §3.3 · p. 8

3.4 Data Collection

• Patient interviews used the same semi-structured protocol across five waves, while practitioner interviews took place only in 2025 and focused on design and decision-making practices.Patient interviews lasted 50–73 minutes and probed concrete episodes, daily use, and long-term behavioral change.Source: §3.4 · p. 8

3.5 Data Analysis

• Trajectory analysis coded 1,182 participant-wave-episodes, and a change counted only when it persisted across waves, was retrospectively narrated, had behavioral evidence, and held across contexts.Three researchers coded the trajectories; practitioner interviews were thematically analyzed to illuminate alignment or mismatch with user practices.Source: §3.5 · p. 8

4.1 The Evolutionary Process of Technology

• The study groups technology change into three accumulating periods, but participants’ adoption timing varied and the periods had no clear boundaries.Some participants encountered mature CGM systems later, without experiencing extended earlier periods.Source: §4.1 · p. 9

4.1.1 Visualization and Continuous Sensing.

• From 2017 to 2019, CGM replaced intermittent fingersticks for many participants and let users view continuous glucose trends on smartphones.Participants described easier measurement and the ability to inspect changing readings and share recent data with a doctor.Source: §4.1.1 · p. 9

4.1.2 Real-Time Feedback and Behavior Triggering.

• From 2019 to 2021, CGM-based real-time feedback included alerts and remote data sharing, while some apps added food tracking and recommendations.Participants described alerts prompting action, family or clinicians noticing abnormalities, and apps linking food entries with glucose data.Source: §4.1.2 · pp. 10–11

4.1.3 Prediction, Recommendation, and Automated Decision-Making.

• From 2021 to 2025, systems began predicting glycemic events and recommending actions, while AID integrated monitoring, insulin delivery, and algorithmic decisions into a continuous loop.One participant said AID reduced the need to calculate before meals, leaving him to confirm approximately.Source: §4.1.3 · p. 11

4.2 User Usage Evolution

• Users’ practices progressed from exploring and interpreting data, to relying on system judgments, and then to system monitoring paired with user response.The paper describes this trajectory as broadly corresponding to successive technology developments.Source: §4.2 · p. 12

4.2.1 Stage 1: Exploration and Sensemaking.

• During exploration and sensemaking, participants compared readings across situations and sometimes cross-checked CGM against fingersticks when values conflicted with bodily sensations.Later adopters encountering more mature feedback technologies tended to show less verification and questioning.Source: §4.2.1 · p. 12

4.2.2 Stage 2: Collaboration and Delegation.

• As feedback became more directive, most users gradually shifted from interpreting data to responding to system conclusions, often experiencing delegated judgment as relief.Ceding remained incomplete: users continued active interactions such as food tracking and data sharing.Source: §4.2.2 · pp. 12–13

4.2.3 Stage 3: Cognitive Disengagement.

• With long-term use, most participants continued wearing devices but increasingly relied on monitoring and acted mainly when an alert or prompt arrived.Information fatigue was one mechanism accelerating reduced engagement; a dangerous AID incident prompted one participant to resume checking data, while another voiced concern about handing everything to machines.Source: §4.2.3 · pp. 13–14

4.3.1 Health Risk.

• Some users increasingly rely on system prompts and may assume no alert means normal, leaving them vulnerable when alerts are not perceived or delivered.P15’s muted device and hearing impairment meant many alerts went unnoticed; the study does not establish how often this pattern causes harm.Source: §4.3.1 · p. 14

4.3.2 Information Load.

• Sustained processing of changing data and prompts creates a continuous burden, and some users respond by checking less or ignoring prompts rather than interpreting data.Users describe fatigue from repeated checking; the resulting pattern is summarized as understanding less while responding more.Source: §4.3.2 · pp. 14–15

4.4 Variation, Subgroup Patterns, and Disconfirming Cases

• Among the 29 participants completing all five waves, 19 (65.5%) broadly followed the three-stage trajectory, while seven (24.1%) showed only the first two stages.Four of the seven were later CGM adopters and may not yet have had time to develop a later-stage trajectory; P9, P22, and P23 resisted the proposed account.Source: §4.4 · p. 15, Figure 1 · p. 16
• Some participants had consistently low interpretive engagement, a stable-low pathway distinct from disengagement, which requires decline from earlier active engagement.The stable-low cases converged on a similar behavioral profile without first developing active interpretation.Source: §4.4 · p. 15
• Withdrawals for unrelated life circumstances were treated as right-censoring, so they do not support inferences about cognitive disengagement.The study conducted brief follow-ups with reachable participants who withdrew, but did not infer trajectories from these withdrawals.Source: §4.4 · pp. 15–16
• Urban and rural completers showed broadly similar trajectories, but rural participants generally reached the stages later as they gained access to technologies later.Higher education was associated descriptively with slower progression and retained verification, but the subgroup was small (n=4) and confounds were uncontrolled.Source: §4.4 · p. 17
• The education pattern is hypothesis-generating rather than a tested moderation effect, given the small higher-education subgroup and lack of controls for adoption depth.The observed comparison cannot separate education from potentially confounding factors.Source: §4.4 · p. 17
• Earlier adopters generally progressed through the stages sooner, while age groups displayed the trajectories with different accounts of letting go or offloading.Older participants used phrases such as “the device is watching for me,” while younger participants described too much information or insufficient time.Source: §4.4 · p. 17

5 Discussion

• Users’ engagement broadly shifts from exploration and sensemaking to collaboration and delegation, and sometimes to continued use with reduced cognitive engagement, but trajectories vary and can reverse.Some participants did not enter the third stage, remained at low engagement, or resumed interpretation after failure or reflection.Source: §5 · p. 17, §5.1 · p. 18

5.1 Three Stages of Changing Cognitive Engagement

• Continued use with cognitive disengagement means devices and data collection persist while self-initiated interpretation declines, unlike abandonment, complacency, or locally measured deskilling.Users engage with data mainly when prompted by an alarm or recommendation; the concept concerns a broader longitudinal reduction in interpretation.Source: §5.1 · pp. 18–19

5.2 The Cost of Cognitive Disengagement

• Clinical prioritization, information compression, and engineering automation jointly favor systems that substitute for judgment, while fatigue and prompt-response burdens accompany reduced independent interpretation.The authors describe health risks when users equate no alarm with normal and a transfer from understanding complex data to continuously responding to prompts.Source: §5.2 · pp. 19–20

5.3 Design Implications: Preserving Interpretive Engagement Under Automation

• Design should automate routine work while keeping consequential reasoning, uncertainty, and context accessible, and preserving meaningful ways for users and care partners to contribute information.The authors also recommend detecting degraded sensing or alert channels and supporting verification and recovery, without requiring continuous manual monitoring.Source: §5.3 · pp. 20–22
• Long-term evaluation should assess whether users can understand consequential outputs, contribute useful context, and recognize degraded conditions, rather than relying on device use or interface activity alone.The proposed design directions are mechanism-informed, not experimentally validated solutions.Source: §5.3 · pp. 21–22

5.4 Generative Implications for Broader Ubiquitous Computing Design

• The findings qualify calm-technology ideals: systems receding into the background may coincide with less frequent independent interpretation rather than assured cognitive empowerment.The discussion presents this as a conceptual lens from diabetes care, not a demonstrated pattern across all ubiquitous systems.Source: §5.4.1 · p. 22

5.4.2 Generative Hypotheses Across Pervasive and Agentic System Classes.

• Possible parallels in automated driving and agentic systems are theoretical extrapolations, not empirically established outcomes outside clinical contexts.The authors present these connections as hypotheses for future investigation.Source: §5.4.2 · p. 23

5.5 Limitations and Future Work

• The 37-person regional cohort is not statistically representative, and the design implications derive from interviews rather than tested interventions.The authors call for cross-context replication and prototyping with longitudinal evaluation.Source: §5.5 · p. 23

6 Conclusion

• Across an eight-year study of 37 patients, continued use with cognitive disengagement involved ongoing device use but less active interpretation, model-building, and questioning.The authors link this state to clinical, designerly, and engineering logics that favor substituting for rather than scaffolding user understanding.Source: §6 · p. 24

• Table 2 is titled “Primary Participant Information.”Source: Table 2 · p. 30

• Table 3 is titled “Practitioner Information.”Source: Table 3 · p. 31

• Table 4 is titled “Interview Guide for Primary Participants.”Source: Table 4 · p. 32

• Table 5 is titled “Interview Guide for Doctors, Designers, and Vendors.”Source: Table 5 · p. 33

Rationale for five waves and two-year intervals

• The eight-year, five-wave design was intended to capture the full trajectory from exploration and sensemaking through delegation to cognitive disengagement across changing technology generations.The two-year intervals balanced time for changes to accumulate, recall accuracy, and cohort feasibility; the authors note some late adopters may need further follow-up.Source: E Detailed Study Design Rationale · p. 33

Rationale for the dual-group design

• Practitioners were interviewed only in 2025 to probe design intent and institutional logic after user-side patterns had emerged, rather than anchor the analysis in practitioner explanations.Their interviews are interpretive, not an independent longitudinal dataset.Source: E Detailed Study Design Rationale · p. 34

Recruitment and sampling strategy

• The 37 patient participants were recruited for variation in residence, education, age, and technology-adoption history, while 15 practitioners were purposively balanced across clinicians, designers, and vendors.Initial recruitment targeted 40 patients to allow for anticipated attrition.Source: E Detailed Study Design Rationale · p. 34

Wave-by-wave technology context

• Technology adoption varied across waves, from glucometers to CGM and then predictive/AID systems, providing within-cohort variation for distinguishing technology effects from time effects.The passage describes these technology contexts as not being experienced simultaneously by all participants.Source: E Detailed Study Design Rationale · p. 34

Interview logistics and setting

• Interviews were conducted one-on-one in participant-chosen private settings, lasted 50–73 minutes, and were recorded with consent and transcribed verbatim within 72 hours.The interviewing researcher verified transcripts against recordings, attending to technical terms and dialectal expressions.Source: F Detailed Data-Collection Procedures · p. 35

Longitudinal consistency protocol

• Patient interviews used the same guide across five waves, with generation-specific additions, and interviewers reviewed prior transcripts to support informed follow-up.Each participant had a primary interviewer for at least three interviews; the remaining interviews were conducted by another researcher to reduce interviewer bias.Source: F Detailed Data-Collection Procedures · p. 35

Interviewer conduct

• Researchers used open, non-evaluative prompts, sought concrete episodes, and offered breaks for discomfort or distress; three participants needed at least one pause in later waves.The participants were P8, P15, and P17.Source: F Detailed Data-Collection Procedures · p. 35

Practitioner interviews

• The 2025 practitioner interviews used a parallel semi-structured protocol focused on user behavior, design or clinical trade-offs, institutional pressures, and gaps between intent and practice.Interviews lasted 45–65 minutes, with a mean of 54 minutes.Source: F Detailed Data-Collection Procedures · p. 35

Participant retention strategy

• Retention efforts included annual check-ins, flexible scheduling, and modest compensation, but 8 of the original 37 participants did not complete all five waves.The authors analyze attrition patterns in §4.4.Source: F Detailed Data-Collection Procedures · pp. 35–36

Data management and security

• Recordings were stored on encrypted drives with access limited to the three coding researchers, and transcripts were de-identified after verification.Data were to be retained for five years after final publication and then securely destroyed.Source: F Detailed Data-Collection Procedures · p. 36

Coding unit rationale

• The participant–wave–episode was the coding unit because interpretive practice was treated as meaningful where context, action, and reasoning co-occur; 1,182 episodes were coded.Episodes ranged from one paragraph to roughly two pages of transcript.Source: G Detailed Data-Analysis Procedures · p. 36

Codebook development and inter-rater reliability

• The stabilized codebook contained 47 codes in six families, and three-way Fleiss’ κ reached 0.81 after dispute resolution on the 246-episode subsample.Initial pairwise κ values ranged from 0.61 to 0.74; the authors treated 0.81 as the reliability floor for the rest of the corpus.Source: G Detailed Data-Analysis Procedures · p. 36

Dispute-resolution protocol

• Disagreements were adjudicated through warrant-sharing and discussion, with a fourth researcher brought in if needed; none of the 246 subsample episodes ultimately required exclusion.Four episodes required adjudication, and 31 reached the discussion step.Source: G Detailed Data-Analysis Procedures · p. 36

Saturation assessment

• Saturation was assessed separately by analytical level and re-evaluated at each wave; descriptive, trajectory, and mechanistic saturation were reached at P21, P26, and P29, respectively.The respective waves were 2021, 2023, and 2025; negative examples were treated as counter-examples and deviant-case data.Source: G Detailed Data-Analysis Procedures · pp. 36–37

Analytic transparency artifacts

• The researchers maintained a versioned codebook, per-participant analytic memos, a disagreement log, and a saturation log for audit and transparency.These artifacts were available to reviewers under a data-sharing agreement.Source: G Detailed Data-Analysis Procedures · p. 37