We conducted 3 semistructured interviews with key stakeholders across the emergency response chain: a paramedic, a fire department unit leader, and a control center dispatcher. All 3 experts were affiliated with the Dortmund Fire Department, which served as the operational partner in the ADLeR (Automatisiertes Detektions-, Melde- und Leitsystem für Rettungskräfte [Automated Detection, Notification, and Guidance System for Emergency Responders]) project. Participants were purposively selected to represent the 3 core operational perspectives along the emergency response process: emergency dispatch and coordination, tactical incident management, and frontline medical care. This maximum-variation approach prioritized structural coverage over the number of interviews. As the interviews served primarily exploratory purposes within the “discover” phase, the resulting themes were subsequently validated and expanded through 2 participatory workshops with 12 additional participants (phase 2).

Following a problem-centered approach [31], interviews explored challenges in alerting, communication, and decision-making under uncertainty. Sessions were conducted using a virtual whiteboard, enabling participants to cocreate digital mind maps that captured workflows, pain points, and preliminary solution ideas. This participatory method aligns with principles of contextual design [32].

Interview data were thematically clustered by 2 researchers (OK and MP) to identify cross-cutting needs and constraints. Each researcher independently reviewed the mind map outputs and assigned thematic labels. Discrepancies were resolved through discussion until consensus was reached. As this step served primarily exploratory purposes, detailed COREQ (Consolidated Criteria for Reporting Qualitative Research) [33] reporting was not applied; however, the interview guide and mind map outputs are provided in Multimedia Appendix 1.

To complement the qualitative findings, descriptive analysis was conducted using publicly available demographic data and anonymized emergency call statistics. The emergency data were provided by the Fire and Rescue Service of the City of Dortmund (Germany) and comprised all medically induced emergency calls recorded in 2021 (n=19,200). The dataset included age group classifications but contained no personally identifiable information. We analyzed age-stratified distributions to identify high-incidence population segments and inform target group selection for subsequent design activities.

The combined findings provided a domain-grounded foundation for phase 2, including identified user needs, operational constraints, and a validated target group profile.

Based on the thematic clusters from expert interviews and demographic indicators from the statistical analysis, we developed 2 representative user personas, capturing attributes such as age, living situation, health status, and technology familiarity. These personas guided the creation of 2 problem scenarios, a residential fire and a fall-related medical emergency, representing prototypical high-risk situations. Each scenario was visualized through low-fidelity storyboards depicting the current emergency response process without technological intervention, highlighting behavioral bottlenecks and communication gaps.

We conducted 2 brainwriting workshops to validate and expand the initial scenarios. Brainwriting is a structured written ideation technique distinct from verbal brainstorming. Participants independently generate ideas in writing before passing them to others for elaboration, which reduces dominance effects and encourages balanced contributions [35]. We applied the 6-3-5 variant, in which 6 participants each write 3 ideas within 5 minutes, then rotate sheets for iterative refinement.

Workshop 1 involved 6 emergency professionals (first responders and dispatch personnel) who reviewed the problem scenarios and contributed design ideas using the 6-3-5 method. The session focused on identifying pain points, desired system behaviors, and integration requirements.

Workshop 2 involved 6 domain experts from academic and applied research institutions with expertise in aging societies, human-machine interaction, and sociotechnical systems. This session addressed system-level considerations, including usability, information interpretability, and edge-case behavior.

Based on workshop outputs, the problem scenarios were iteratively transformed into activity scenarios incorporating envisioned system functionalities. These were synthesized into a low-fidelity prototype combining key system elements into a coherent, user-facing concept.

We conducted workshop sessions online using digital whiteboards. Exported materials served as the primary qualitative data source. Two researchers (AR and MP) independently analyzed the content following qualitative content analysis principles, inductively coding idea clusters to identify recurring patterns and design requirements. Discrepancies were resolved through iterative discussion until consensus was reached. Reporting follows the COREQ guideline [33]; details are provided in Multimedia Appendix 2.

Validated activity scenarios, annotated storyboards, and a low-fidelity prototype serving as stimulus material for the acceptance evaluation.

Survey constructs were based on the Technology Usage Inventory (TUI) [36], which extends traditional acceptance models by incorporating individual traits including curiosity (NEU), anxiety (ANG), interest (INT), perceived usefulness (NÜT), user-friendliness (BEN), skepticism (SKE), and accessibility (ZUG). Abbreviations reflect the original German instrument. In the TUI framework, accessibility refers to the perceived availability of and ease of access to the technology, including how readily it can be obtained, set up, and integrated into daily routines. This construct does not refer to disability-related accessibility in the sense of barrier-free design. Two additional validated measures were included: the social compatibility (SOC) scale [37] for prosocial attitudes and the activities of daily living (ADL) scale [38] for functional independence.

The survey used a pre-post design: participants first completed demographic questions, smart home experience items, and baseline constructs (NEU, ANG, ADL, and SOC). They were then exposed to visual storyboards illustrating the emergency scenarios with integrated smart home functions (developed in phase 2). Post-exposure, participants rated the remaining TUI constructs and intention to use (ITU). This design enabled measurement of both general dispositions and concept-specific evaluations.

A pilot test (n=5) ensured item clarity and technical robustness. Survey implementation followed CHERRIES (Checklist for Reporting Results of Internet E-Surveys) reporting standards [39]; details are provided in Multimedia Appendix 3.

The survey was implemented as an open online questionnaire using LimeSurvey, hosted on a secure university server. It was distributed between October 2022 and June 2023 through multiple channels: housing cooperatives and senior advisory councils served as the primary recruitment partners, and the Dortmund Fire Department shared the survey link via social media (X, Facebook, and LinkedIn). Participants were also able to share the link independently. No financial or material incentives were offered. The only inclusion criterion was the ability to complete the German-language questionnaire; no further restrictions regarding age, health status, or technology experience were applied. A total of 128 individuals accessed the survey, of whom 85 completed all 4 sections (completion rate: 66%). Cookies were used to prevent repeated participation from the same device. The final sample included adults aged 20 to 89 years.

We performed analyses in R (version 4.3.3) using lavaan (version 0.6‐20). We evaluated internal consistency using Cronbach α, with values ≥0.70 considered acceptable [40]. A 2-stage modeling strategy combined confirmatory factor analysis (CFA) and structural equation modeling (SEM). Given ordered-categorical measurement and the number of indicators per construct, a diagonally weighted least squares (DWLS) estimator with robust corrections was used. We evaluated model fit using scaled indices (chi-square, comparative fit index [CFI], Tucker-Lewis index [TLI], root mean square error of approximation [RMSEA], and standardized root mean square residual [SRMR]), with CFI/TLI >0.90 and RMSEA <0.08 interpreted as acceptable.

While no universally accepted minimum sample size exists for DWLS estimation, simulation research suggests that DWLS with robust corrections can yield adequate parameter estimates and fit indices at sample sizes below 200, particularly when indicators are ordinal and factor loadings are moderate to strong [41,42]. Nevertheless, with 85 participants and 9 latent predictors, the ratio of sample size to model complexity is limited, which may reduce statistical power for detecting moderate effects and increase the uncertainty of individual path estimates. The results should therefore be interpreted with appropriate caution regarding the precision of nonsignificant coefficients.

Validated acceptance model identifying key facilitators and barriers to adoption intention.

Across the 3 interviews conducted, a total of 82 individual idea nodes were cocreated using digital whiteboarding tools, resulting in 3 structured mind maps (Multimedia Appendix 1). Thematic clustering revealed recurring barriers and requirements along 3 operational layers: dispatch coordination, responder navigation, and lay responder integration.

A central theme concerned the limited situational awareness available to dispatchers at the time of call intake. Interviewees described frequent uncertainty due to missing contextual data, such as the caller’s precise location, environmental hazards, or patient-specific health risks.

First responders, in turn, reported considerable time losses caused by restricted access to buildings and unclear apartment layouts. Navigating unfamiliar environments, particularly under time pressure, was cited as a major operational burden. Participants noted that prior knowledge about access points, apartment configurations, or mobility limitations could significantly streamline response efforts.

Finally, the role of lay responders proved both promising and problematic. While early arrival can be beneficial, the lack of standardized guidance and limited integration into formal rescue coordination restricts their effectiveness. Suggestions included structured onboarding, risk-adapted task delegation, and access to context-specific information prior to arrival.

Descriptive analysis of age-stratified emergency statistics confirmed the relevance of older adults as a disproportionately affected population segment. Although individuals aged 60 years and above represent less than one-third of the general population, they accounted for 64% (12,232/19,200) of all medically induced emergency calls in 2021 (Table 2). This disproportionality indicates that age is not merely a proxy for population size but a significant factor in emergency system usage. Thus, older adults were selected as the primary target group for subsequent design phases.

These findings emphasize the need for improved information availability, environmental transparency, and support mechanisms for nonprofessional actors, which directly informed the scenario development in phase 2.

The structured idea generation yielded a diverse set of system-relevant concepts, which were subsequently clustered into key design implications. Participants emphasized the importance of real-time contextual awareness enabled by ambient and body-worn sensor data, precise geolocation, and adaptive alarm routing tailored to user roles and incident types. Suggestions also included infrastructure-level support for building access and orientation, such as dynamic lighting, illuminated door signage, and smart locking mechanisms.

Several contributions addressed systemic integration challenges, highlighting the need for interoperability across heterogeneous devices and platforms. The envisioned system was expected to balance automation with user override options and to provide differentiated access to sensitive data, depending on responder qualification and proximity. Participants also raised concerns regarding privacy, false-positive prevention, and reliability under edge-case conditions.

Based on these inputs, we revised the initial scenarios into detailed activity scenarios incorporating technical functionalities and user interactions, visualized through annotated low-fidelity storyboards.

A subsequent workshop evaluated these scenarios, focusing on usability, privacy, and error tolerance. Feedback was synthesized into a low-fidelity prototype illustrating key components and user touchpoints.

Figure 2 illustrates selected panels from the fire emergency scenario, depicting the sequence from unnoticed hazard through autonomous detection, system-initiated intervention, and successful responder arrival (see Multimedia Appendix 2 for complete storyboards).

The resulting artifacts, including 29 design requirements and annotated storyboards (Multimedia Appendix 2), served as the conceptual foundation for the acceptance study conducted in phase 3.

A total of 85 participants completed all 4 sections of the survey. The sample included individuals aged 20 to 89 years, with the most represented groups being those aged 30‐39 years (n=23, 27%) and 20‐29 years (n=16, 19%). A majority of participants identified as male (n=55, 65%), and most lived either in one-person (n=28, 33%) or two-person households (n=36, 42%). Regarding education, 44% (n=37) reported a college degree, while 29% (n=25) held a secondary school qualification.

Current smart home usage was reported by 61% (52/85) of participants, with 33% (28/85) indicating an intention to adopt such technologies within the next 12 months. Adoption intentions varied slightly by age, gender, and living arrangement (see Table 3).

Descriptive statistics for all psychometric constructs and individual items are summarized in Table 4. Participants reported high levels of functional independence, with mean scores above 4.6 across all ADL items (eg, ADL1: mean 4.81, SD 0.72). Social compatibility showed moderate to high values (SOC1-SOC5: mean>4.2), though lower agreement was observed for items reflecting burdensome social responsibility (SOC6-SOC9: mean≈3‐4).

Pre-exposure assessments revealed high curiosity (NEU1: mean 6.04, SD 1.19) and low technology-related anxiety (ANG1: mean 2.38, SD 1.58). Following the storyboard intervention, participants rated the proposed system as generally useful (NÜT1: mean 5.99, SD 1.11) and user-friendly (BEN1: mean 5.20, SD 1.35), with only moderate skepticism. Accessibility received more varied evaluations (ZUG1: mean 3.60, SD 1.49).

We measured the ITU via a 0‐100 visual analog scale. Participants reported moderately high usage intentions, with item means ranging from 70.34 (SD 23.98) to 81.59 (SD 17.95) across the three ITU items.

Before conducting the latent-variable analyses, we performed exploratory comparisons to examine whether ITU differed across sociodemographic and contextual subgroups. Figure 3 displays the distribution of ITU scores by age group, education level, gender, living situation, and smart-home–related characteristics (ie, presence of smart-home devices in the household and planned future purchases).

Across age categories (Figure 3A), median ITU scores appeared largely comparable, with slightly higher intentions in the youngest (20‐29 years) and older (70‐79 years) participants, but no consistent age-related trend. Regarding smart-home–related factors, participants who already had smart-home devices in their household (Figure 3E) tended to report slightly higher ITU scores than those without such devices, although the spread of values was wide and overlapping. Similarly, those who planned to purchase smart-home technologies in the near future (Figure 3F) reported somewhat elevated intentions relative to those without purchase plans, but again without clear statistical separation. Given the absence of significant mean-level differences in ITU across sociodemographic subgroups, these variables were not included as exogenous predictors in the SEM.

Overall, these analyses indicated no meaningful sociodemographic bias in ITU across the examined groups. This supports the focus on psychological and attitudinal constructs as the main determinants in the subsequent latent-variable modeling.

Before estimating the latent variable model, we examined reliability and zero-order associations among the observed constructs. All multi-item scales demonstrated good internal consistency, with Cronbach α values ranging from 0.72 (BEN) to 0.93 (ADL) (see Table 5).

Bivariate correlations showed that NÜT (r=0.60, P<.001), ZUG (r=0.44, P<.001), SOC (r=0.37, P=.001), NEU (r=0.37, P=.001), INT (r=0.24, P=.03), and BEN (r=0.24, P=.03) each correlated positively with ITU. In contrast, SKE correlated negatively (r=–0.54, P<.001), while ANG (r=–0.11, P=.31) and ADL (r=–0.06, P=.56) showed small, nonsignificant associations.

Simple linear regressions mirrored these bivariate patterns: positive effects on ITU emerged for NÜT, ZUG, SOC, NEU, INT, and BEN (all P≤.03), whereas SKE exerted a significant negative effect (β=–6.82, P<.001). Neither ANG nor ADL reached statistical significance (P>.30).

Together, these findings supported the proposed directionality of the theoretical model and justified proceeding to latent-variable analyses.

We applied a comprehensive confirmatory factor and SEM approach using a DWLS estimator with robust corrections, appropriate for ordered-categorical indicators and the comparatively high number of observed variables per construct. The 9-factor measurement model, plus a 3-indicator factor for ITU, demonstrated an acceptable overall model fit on scaled indices (χ²₈₁₅=1052.21, P<.001; CFI=0.928; TLI=0.920; RMSEA=0.059, 90% CI 0.048‐0.069; SRMR=0.107; Bentler-SRMR=0.0923) (Table 6).

Although both SRMR indices were slightly above conventional thresholds [43], simulation research indicates that SRMR tends to be inflated when using categorical estimators such as DWLS, particularly in complex models with many indicators [44]. The combination of CFI and TLI values close to 0.93 and an RMSEA below 0.06, given the categorical measurement level and the model’s complexity, supports an adequate and stable representation of the empirical covariance structure. All standardized factor loadings were positive, substantial, and statistically significant (P<.001), ranging from 0.46 to 0.99, with most exceeding 0.70. No cross-loadings or Heywood cases were detected, and the factor solution remained stable across estimation runs, confirming the stability of the measurement specification.

The SEM maintained an acceptable overall fit that closely mirrored the CFA solution and accounted for 66.1% of the variance in ITU (R²=0.661). Among the latent predictors, ZUG was the strongest positive determinant (β=0.325, P=.02). In contrast, SKE showed a negative association with ITU that approached but did not reach conventional significance levels (β=−0.336, P=.06). NÜT (β=0.357, P=.15) and SOC (β=0.129, P=.25) showed positive but nonsignificant trends, while BEN, INT, NEU, ANG, and ADL displayed no meaningful associations (|β|≤0.26, all P>.10; Table 7). A comparison between the full model (with all 9 predictors) and a reduced model (retaining only significant predictors) revealed minimal differences in fit indices (ΔCFI=−0.003; ΔRMSEA=0.004; ΔSRMR=0.004). Although a nested model test indicated a statistically significant decrement in fit (P=.001), the negligible changes in practical fit indices suggest that retaining theoretically motivated but statistically nonsignificant constructs did not materially distort the overall model.

The final structural equation model is depicted in Figure 4, illustrating the pattern of relationships among the 9 latent predictors and ITU. All exogenous latent constructs were allowed to covary, controlling for shared variance among psychological, functional, and social factors. These correlations are depicted as curved double-headed arrows in the structural model but were not interpreted further, as they were not part of the primary hypotheses.

To assess the robustness of the structural findings, we conducted several complementary diagnostics. To ensure that the constructs included in the latent-variable model were empirically distinct and not affected by multicollinearity, we performed additional post hoc diagnostics. The latent factor correlations extracted from the CFA model indicated satisfactory discriminant validity, as none of the interfactor correlations approached unity and all remained well below critical thresholds (|r|<0.85). This supports the conceptual distinctness of the modeled constructs.

Variance inflation factors (VIF) confirmed the absence of problematic multicollinearity, with values ranging from 1.24 to 1.81 (all VIF<5), indicating that the predictors were sufficiently independent. Alternative model specifications (eg, method factors, item parceling) provided no meaningful improvement and were not retained.

The findings provide a coherent picture of the psychological and contextual factors underlying intention to use. The next section discusses their theoretical and practical implications.