Accurately estimating energy expenditure (EE) is crucial for understanding exercise efficiency, managing fitness goals, and monitoring health conditions. Existing wearable systems either rely primarily on heart rate and motion sensors, resulting in unsatisfactory accuracy, or require bulky setups such as thermal cameras to improve performance by integrating more physiological information, which limits real-world applicability. We present EarCalo, an earable-based system that leverages in-ear audio sensing to estimate EE during running. The system extracts airflow-induced acoustic variations within the ear canal and employs a deep neural network to translate these subtle in-ear sound dynamics into EE estimates. The key insight is that in-ear acoustic signals can capture multiple physiological factors such as breathing and cardiovascular activity, while also reflecting motion-related cues like running intensity. These rich acoustic cues are closely related to EE and provide a unified sensing modality for estimation. We evaluated EarCalo on 21 participants running at varying speeds using a mixed-user setting. EarCalo achieved a mean absolute error (MAE) of 0.67 kcal/min, a mean absolute percentage error (MAPE) of 11.98%, and a Pearson correlation of 0.945, which achieves accuracy close to established physiological standards. This work represents an early step toward practical and personalized earable-based EE estimation in everyday settings.

Background About 14.9% of people (9.9 million) registered with primary care practices in England and Wales are prescribed medication for hypertension. However, many do not take their medication as prescribed. To address this problem, we need scalable interventions. Objective To develop a scalable low-cost intervention to support medication adherence in people prescribed medication for hypertension in primary care, and to obtain precise and robust estimates of the effectiveness and cost-effectiveness of the intervention compared with usual care. Design Systematic reviews and meta-analyses; qualitative meta-synthesis; interviews and focus groups; expert consultations; pre-testing study; randomised feasibility trial; randomised controlled trial of effectiveness and cost-effectiveness; economic modelling. Setting and participants Primary care practices in England and Wales. Patients prescribed medication for hypertension with poorly controlled blood pressure. Interventions Very brief intervention delivered by a practice nurse or healthcare assistant followed by a digital intervention (text messaging programme or smartphone app). Main outcome measures Acceptability, feasibility, fidelity and cost of the interventions. Systolic blood pressure. Biochemical and self-reported measures of medication adherence. Results Our systematic reviews showed that both app-based and face-to-face interventions in patients with long-term conditions have a positive effect on medication adherence. The meta-synthesis of published qualitative studies showed that: digital interventions to support medication use were perceived as acceptable and useful; a digital intervention would be more effective if it was personalised and tailored; barriers to using digital interventions included lack of interest, lack of confidence and lack of proficiency and experience in using the technology; digital interventions should be simple, easy to navigate and age-appropriate; patients wanted accurate information on their health condition, potential side effects of medication and health consequences of non-adherence; reminder notifications and a self-monitoring feature were perceived as helpful by some patients; some patients suggested that a digital intervention should enable them to communicate with pharmacies, but practitioners were concerned that this would increase their workload. The interview and focus group study identified several barriers to adherence, including forgetting, unpleasant side effects and reluctance to medicate. A digital intervention to support medication adherence was acceptable to patients if it was user-friendly, the content was tailored to the user, and the privacy of user data was protected. Simple reminder messages for taking medication and reordering prescriptions were considered more useful by patients than those providing information on the benefits of medication or the consequences of non-adherence. Patients preferred to receive feedback on their adherence levels in the form of a simple graph, percentage score or statistic. Practitioners thought that it would be feasible to introduce a digital intervention to patients in a very brief face-to-face consultation. In the pre-testing study, participants reported that the interventions we developed were easy to use and that they would recommend them to others. The feasibility trial showed that the combined intervention was acceptable and that a large cost-effectiveness trial was feasible. The main trial showed no difference between arms in systolic blood pressure or medication adherence at 12-month follow-up. The estimate (95% confidence interval) for the difference in means between arms in self-measured systolic blood pressure at 12 months was −0.61 mmHg (−3.05 to 1.82), p = 0.62 [for intervention vs. control (reference group)]. In the base case analysis, the intervention had a mean incremental cost-effectiveness ratio below the usual willingness-to-pay thresholds in the National Health Service in the United Kingdom. The probability of cost-effectiveness was between 77% and 80% at willingness-to-pay thresholds of £15,000, £20,000 and £30,000 per quality-adjusted life-year, but the confidence intervals are wide and cross zero, indicating some chance that the intervention could be less effective and more costly. Limitations The effectiveness trial was conducted during the COVID-19 pandemic. To reduce the risk of infection, the very brief intervention was delivered by telephone instead of face-to-face, and all study measurements were conducted remotely. This may have led to lower response rates and data quality and lower effectiveness of the intervention. A significant proportion (24%) of participants did not have raised blood pressure at baseline, and self-reported medication adherence was high at baseline, which reduced the possible scope for an intervention effect. Conclusions The findings on effectiveness do not support the commissioning of the intervention in United Kingdom primary care. The cost-effectiveness findings are more equivocal, showing a high probability of being cost-effective at standard United Kingdom willingness-to-pay thresholds, but with some uncertainty. Future work Future research should address the challenge of identifying and recruiting people who are poorly adherent and have raised blood pressure and test the intervention in this group. Variants such as face-to-face delivery, adding a follow-up consultation or a purely digital version could also be investigated. Study registration This study is registered as CRD42017080150; CRD42020164049; ISRCTN12805654; ISRCTN74504989; ISRCTN82013652. Funding This award was funded by the National Institute for Health and Care Research (NIHR) Programme Grants for Applied Research Programme (NIHR award ref: RP-PG-0615-20013) and is published in full in Programme Grants for Applied Research; Vol. 14, No. 18. See the NIHR Funding and Awards website for further award information. Plain language summary About 15% of people (10 million) registered with primary care practices in England and Wales are prescribed medication for hypertension (high blood pressure). However, many do not take their medication as prescribed, which may harm their health and increases health service costs. To address this problem, we need low-cost interventions that can reach all the people who need them. This research programme aimed to develop a new intervention to support people with poorly controlled blood pressure to take their hypertension medication as prescribed and to assess how well it worked and how much it cost compared with usual care only. The findings would inform a decision on whether to introduce the intervention in primary care practices in the United Kingdom. We used a range of different research methods, including literature reviews, interviews and focus groups, and randomised controlled trials. The intervention we developed has two parts: a very brief intervention delivered by a practice nurse or healthcare assistant by telephone, followed by a digital intervention (individually tailored text messages for up to 420 days or a smartphone app). In the main trial, 537 participants received either the intervention or just usual care and were followed up 12 months later. The intervention was low cost, but it was not effective: the results showed no difference in blood pressure or medication adherence between the two groups at 12 months. Not all participants had raised blood pressure at the start of the study, and many were already taking their medication as prescribed; this may explain why the intervention did not work. The findings suggest that the intervention should not be introduced in United Kingdom primary care practices before showing that it works in patients who do not take their medication as prescribed and have raised blood pressure. Scientific summary Background Hypertension is a global health challenge accounting for 8.5 million deaths worldwide despite the availability of low-cost pharmaceutical treatment. About 14.9% of people (9.9 million) registered with primary care practices in England and Wales are prescribed medication for hypertension. However, many patients with hypertension do not take their medications as prescribed – 42% in the UK according to one study. Non-adherence to antihypertensive medication is associated with increased risk of suboptimal blood pressure (BP) control, complications and all-cause mortality, and increased healthcare costs. Primary care practitioners have an important role in supporting patients to adhere to their prescribed medication. However, they lack time to provide ongoing support for adherence, and their time is expensive. A potential solution is for a practitioner such as a practice nurse to deliver a very brief intervention (VBI) during a consultation and to use a digital intervention such as text messaging or a smartphone app to support subsequent adherence. About 96% of UK adults use a mobile phone, and in 93% of cases, this is a smartphone; the corresponding figures for those aged 65 and above are 88% and 77%. This suggests that digital interventions have the potential to reach the majority of this population. Digital interventions have several other advantages over traditional interventions: they can be fully automated; provide information that is highly tailored to the individual; be interactive; be available at any time; deliver support in real time; deliver support with high fidelity; and be easily updated. Recent meta-analyses have reported promising findings for the effectiveness of nurse-led and digital interventions to improve medication adherence and reduce BP in people with hypertension. The interventions examined in these reviews varied widely in content and delivery, and the digital interventions that have been evaluated to date have not made full use of individual tailoring, interactivity and other features that may increase user engagement and potential effectiveness. Objectives The PAM programme (Programme on Adherence to Medication) aimed to develop and evaluate an intervention to support medication adherence that combines a VBI from a practice nurse or healthcare assistant with a digital intervention (text messaging programme or smartphone app). Such an intervention would be inexpensive to deliver, scalable and potentially cost-effective. The objectives were: To develop a scalable low-cost intervention to support medication adherence in people prescribed treatment for hypertension in primary care. To evaluate the acceptability and feasibility of the intervention and the feasibility of conducting a (cost-)effectiveness trial. To provide precise and robust estimates of the effectiveness and cost-effectiveness of the intervention compared with usual care (UC). To develop an economic model of the cost-effectiveness of medication adherence interventions. To inform a decision on whether to implement the intervention in primary care. Methods The target group is patients in primary care practices in England and Wales who do not take their antihypertensive medication as prescribed and have raised BP. Methods used were: Systematic reviews of randomised controlled trials of app-based (9 trials) and face-to-face interventions (20 trials) to support medication adherence, with random-effects meta-analyses. A meta-synthesis of 30 published qualitative studies of adults taking medication for cardiovascular-related long-term health conditions (e.g. type 2 diabetes, hypertension) and/or healthcare practitioners who treat patients with cardiovascular conditions who were asked about their views and experiences of digital interventions to support medication adherence. Interviews with 11 healthcare practitioners [6 practice nurses, 2 healthcare assistants, 2 practice pharmacists, 1 general practitioner (GP)] and 6 patients, and 4 focus groups, with a total of 14 patients, to gather views on the acceptability and content of digital interventions for medication adherence. Expert consultations with 2 commissioners, 2 academics, 2 nurses, 2 patients and 10 GPs. Participants were e-mailed a description of the proposed intervention, a description of the proposed design of the randomised feasibility study and a link to an online questionnaire that asked their views on the delivery mode and content of the intervention and on the proposed feasibility study. Commissioners were asked about the evidence needed to inform a decision whether to commission the intervention. Pre-testing study of early versions of the digital interventions to assess acceptability. The text messaging intervention was used by 22 patients with hypertension for 28 days, and four of them also used the smartphone app for an additional 28 days. Data were collected by weekly telephone interviews, questionnaires and log files showing how they had used the interventions. Randomised feasibility trial to assess the feasibility and acceptability of the PAM intervention and the feasibility of conducting a large cost-effectiveness trial. Patients with hypertension who had raised BP and were non-adherent to their prescribed medication as indicated by their practice records and practice GP assessment were eligible for the study. One hundred and one eligible patients from nine general practices in the East of England and London were randomised to receiving the PAM intervention (N = 61) or UC only (N = 40). Randomised controlled trial to estimate the effectiveness and cost-effectiveness of the PAM intervention to improve medication adherence and reduce BP compared with UC only, to inform a decision on whether to implement the intervention in primary care (‘main trial’). A total of 573 eligible patients from 57 practices in England and Wales were individually randomised, stratified by practitioner, to the PAM intervention or control (UC only) and followed up at 12 months. The primary outcome was systolic blood pressure (SBP). The analysis was based on 537 participants. Within-trial economic analysis of the cost-effectiveness of the PAM intervention compared with UC alone. The main cost-effectiveness measure was the incremental cost per quality-adjusted life-year (QALY) gained, and the analysis included extensive deterministic and probabilistic sensitivity analyses. Results Systematic reviews The findings from the meta-analysis of app-based interventions showed that, at follow-up, patients in the intervention groups were more likely to self-report adherence to medication than those in the comparator groups [odds ratio 2.12, 95% confidence interval (CI) 1.64 to 2.75, n = 988, p < 0.0005]. None of the behaviour change techniques (BCTs) used in the interventions was significantly associated with intervention effect size. In the meta-analysis of face-to-face interventions, statistically significant pooled effects were found favouring the intervention arm over the control arm for several Medication Event Monitoring System measures of adherence, for example, percentage of prescribed doses taken on time over a period of 3 weeks to 2 months [mean difference (MD) 9.34, 95% CI 4.36 to 14.33, n = 3,667, p = 0.0002]. We also found significant between-arm effects for a self-report measure of adherence (Morisky scale). The impact of BCTs on intervention effectiveness could not be estimated as the analyses were underpowered. Taken together, these reviews supported our proposal to use face-to-face and digital components in the PAM intervention. However, we were unable to identify promising BCTs for potential inclusion in the proposed intervention. Meta-synthesis of previous qualitative studies The main findings from the meta-synthesis of published qualitative studies were: digital interventions to support medication use were perceived as acceptable and useful; a digital intervention would be more effective if it was personalised and tailored; barriers to using digital interventions included lack of interest, lack of confidence and lack of proficiency and experience in using the technology; digital interventions should be simple, easy to navigate and age-appropriate; patients wanted accurate information on their health condition, potential side effects of medication and health consequences of non-adherence; reminder notifications and a self-monitoring feature were perceived as helpful by some patients but unnecessary by others; some patients suggested that a digital intervention should enable them to communicate with pharmacies, but practitioners were concerned that this would increase their workload. Interviews and focus groups with practitioners and patients This study identified several barriers to adherence, including forgetting, unpleasant side effects and reluctance to medicate. A digital intervention to support medication adherence, either via text messages or smartphone app, was acceptable to patients, provided that it was user-friendly, the content was tailored to the user and the privacy of user data was protected. Simple reminder messages for taking medication and reordering prescriptions were considered more useful by patients than those providing information on the benefits of medication or the consequences of non-adherence, which were favoured by practitioners. Rather than messages of encouragement, patients preferred to receive feedback on their adherence levels in the form of a simple graph, percentage score or statistic. All the practitioners thought that it would be feasible to introduce a digital intervention to patients in a very brief face-to-face discussion during a primary care consultation. Expert consultations There was substantial similarity of views between the different stakeholders. They found the concept of a VBI delivered face-to-face by a healthcare practitioner acceptable. However, they felt that it was not feasible to address possible reasons for medication non-adherence in a VBI, and that the intervention should be limited to emphasising the importance of taking medication as prescribed and signposting the patient to a digital intervention. Pre-testing study Participants reported that the interventions were easy to use and that they would recommend them to other people. They were satisfied with the frequency of the messages and the content of the daily reminder and weekly query messages, but they were somewhat less satisfied with the content of the daily non-reminder (advice) messages. The response rate to the query messages was 100%, indicating a high degree of engagement. Randomised feasibility trial All 101 participants had their BP measured at baseline, and the vast majority provided a urine sample for chemical adherence testing. Participants were on average 65.8 years of age, 54% male, with a substantial minority (35%) from the most deprived areas, based on practice postcode. Baseline characteristics were similar in the two arms. At 3-month follow-up, 83% of participants had their BP measured and provided a urine sample, and the percentage was similar in the two arms. Ninety-two per cent of participants randomised to the intervention arm opted to receive text messages, and 8% opted to use the app. Ninety per cent responded to the tailoring questions which were administered digitally. Four intervention participants actively disengaged from the digital intervention by sending a STOP message. Seventy-two per cent continued to use the digital intervention for at least 1 month. The post-trial interviews showed that intervention participants found the intervention to be acceptable. Participants were satisfied with the baseline and follow-up consultations and the study procedures, and there were no concerns among control participants about being randomised to this arm. Practitioners also confirmed that the study procedures and intervention were acceptable. From baseline to follow-up, mean SBP reduced from 146.9 mmHg to 136.9 mmHg in the intervention arm compared with no change in the control arm (adjusted MD 9.2 mmHg, 95% CI 5.7 to 12.6), and biochemically measured adherence increased to a greater extent in the intervention arm than in the control arm, suggesting that the intervention was potentially effective. The findings from this trial showed that the intervention was acceptable to participants and that most offered the digital intervention used it, at least in the short term. The trial procedures were demonstrated to be practicable. Together with the findings on trial uptake and retention rates, this suggested that a large cost-effectiveness trial was feasible. Main trial Baseline characteristics were similar in the two arms. The majority of participants were recruited from practices in the East of England. Similar to the feasibility trial, 56% were male and mean age was 66.5 years. The vast majority categorised themselves as being of White ethnicity, but there was a range of deprivation levels, based on participant home postcode. Mean BP, obtained from practice records before randomisation, was 145/82 mmHg, again similar to the sample in the feasibility trial. Of participants, 75.8% had a BP reading above the accepted cut-off of 140/90 mmHg and 71.7% had a SBP reading above 140. We were, therefore, partially successful in recruiting a sample of participants who had a raised BP even though they were prescribed antihypertensive medication. The estimate (95% CI) for the difference in means between arms in the primary outcome of self-measured SBP at 12 months was −0.61 mmHg (−3.05 to 1.82), p = 0.62 [for intervention vs. control (reference group)]. Thus, the estimated effect was very small, and the detectable effect size of 5 mmHg did not fall within the CI. The estimate for the difference in means between arms for SBP obtained from practice records was 1.04 mmHg (−1.56 to 3.65), p = 0.43. Thus, the intervention appeared to have no effect on SBP. There was also no effect of the intervention on biochemically measured adherence. Of the 392 participants who provided a urine sample at follow-up, 388 (99.0%) were found to have at least one antihypertensive medication (or metabolite) in their urine. Based on the urinalysis, 94.3% of participants were categorised as ‘fully adherent’, 4.8% as ‘partially adherent’ and only 0.9% as ‘non-adherent’. There was no difference in these percentages between trial arms. Self-reported adherence at 12 months was also high, with no difference between trial arms. The mean score on the five-item Medication Adherence Report Scale questionnaire was 23.8 [standard deviation 1.7] out of a maximum score of 25 (based on 387 participants who provided 12-month data on this scale). Of the intervention participants, 212 (77.9%) opted to receive text messages; 60 (22.1%) opted to use the app, and 39 of these became active users. On average, participants were satisfied with the combined intervention (VBI plus digital intervention) and thought that it was acceptable and effective. However, 61.0% used the digital intervention for less than 3 months, with the main reasons being not needing any further support and finding the messages annoying. Economic analysis The total mean intervention cost per patient was £30. In the base case analysis, the intervention was found to be cost-effective compared with UC, with a mean estimated incremental cost-effectiveness ratio (ICER) of £1231 per QALY gained (95% CI −£13,156 to £19,535) and mean incremental net monetary benefit (INMB) of £289 (95% CI −£498 to £1026) at a willingness-to-pay (WTP) threshold of £15,000/QALY. The INMB rose to £400 (95% CI −£643 to £1383) and £621 (95% CI −£933 to £2104) for £20,000/QALY and £30,000/QALY, respectively. The probability that the intervention is cost-effective was between 77% and 80% for these WTP thresholds. Limitations The effectiveness trial was conducted during the COVID-19 pandemic. To reduce the risk of infection, the VBI was delivered by telephone instead of face-to-face, and all study measurements were conducted remotely. This may have led to lower response rates and data quality and lower effectiveness of the intervention. A significant proportion (24%) of participants did not have raised BP at baseline, and self-reported medication adherence was high at baseline, which reduced the possible scope for an intervention effect. Conclusions The combination of a VBI delivered by a practice nurse or healthcare assistant and a digital intervention was acceptable to both patients and practitioners. However, although the feasibility trial showed promising results, the main trial showed no effect of the intervention on medication adherence or SBP at 12 months. The cost-effectiveness findings showed a mean ICER below the usual WTP thresholds, but the CIs are wide and cross zero, indicating some chance that the intervention could be less effective and more costly compared with UC only. The effectiveness findings do not support the commissioning of the intervention in UK primary care. Future research should address the challenge of identifying and recruiting people who are poorly adherent. If it is feasible to recruit patients who are non-adherent to their prescribed antihypertensive medication and have raised BP, the intervention could be tested with the VBI delivered remotely or face-to-face. Variants such as adding a follow-up consultation or testing a purely digital version could also be investigated. The economic model developed for this programme can provide the basis for future economic evaluations of similar interventions across a range of different conditions. Study registration This study is registered as CRD42017080150; CRD42020164049; ISRCTN12805654; ISRCTN74504989; ISRCTN82013652. Funding This award was funded by the National Institute for Health and Care Research (NIHR) Programme Grants for Applied Research Programme (NIHR award ref: RP-PG-0615-20013) and is published in full in Programme Grants for Applied Research; Vol. 14, No. 18. See the NIHR Funding and Awards website for further award information.

Time series data from the Intensive Care Unit (ICU) provides critical information for patient monitoring. While recent advancements in applying Large Language Models (LLMs) to time series modeling (TSM) have shown great promise, their effectiveness on the irregular ICU data, characterized by particularly high rates of missing values, remains largely unexplored. This work investigates two key components underlying the success of LLMs for TSM: the time series encoder and the multimodal alignment strategy. To this end, we establish a systematic testbed to evaluate their impact across various state-of-the-art LLM-based methods on benchmark ICU datasets against strong supervised and self-supervised baselines. Results reveal that the encoder design is more critical than the alignment strategy. Encoders that explicitly model irregularity achieve substantial performance gains, yielding an average AUPRC increase of 12.8% over the vanilla Transformer. While less impactful, the alignment strategy is also noteworthy, with the best-performing semantically rich, fusion-based strategy achieving a modest 2.9% improvement over cross-attention. However, LLM-based methods require at least 10× longer training than the best-performing irregular supervised models, while delivering only comparable performance. They also underperform in data-scarce few-shot learning settings. These findings highlight both the promise and current limitations of LLMs for irregular ICU time series. The code is available at https://github.com/mHealthUnimelb/LLMTS.

Wearable foundation models (WFMs), trained on large volumes of data collected by affordable, always-on devices, have demonstrated strong performance on short-term, well-defined health monitoring tasks, including activity recognition, fitness tracking, and cardiovascular signal assessment. However, most existing WFMs primarily map short temporal windows to predefined labels via static encoders, emphasizing retrospective prediction rather than reasoning over evolving personal history, context, and future risk trajectories. As a result, they are poorly suited for modeling chronic, progressive, or episodic health conditions that unfold over weeks, months or years. Hence, we argue that WFMs must move beyond static encoders and be explicitly designed for longitudinal, anticipatory health reasoning. We identify three foundational shifts required to enable this transition: (1) Structurally rich data, which goes beyond isolated datasets or outcome-conditioned collection to integrated multimodal, long-term personal trajectories, and contextual metadata, ideally supported by open and interoperable data ecosystems; (2) Longitudinal-aware multimodal modeling, which prioritizes long-context inference, temporal abstraction, and personalization over cross-sectional or population-level prediction; and (3) Agentic inference systems, which move beyond static prediction to support planning, decision-making, and clinically grounded intervention under uncertainty. Together, these shifts reframe wearable health monitoring from retrospective signal interpretation toward continuous, anticipatory, and human-aligned health support.

Speculative decoding accelerates LLM inference by drafting candidate tokens and verifying them in parallel. Tree-attention drafters such as EAGLE-3 are widely adopted, yet typically hold two decisions fixed: (1) a strict token-match verification rule and (2) a static draft-tree shape. Prior work relaxes each in isolation under limiting assumptions: long draft chains for training-free lossy verification, and adaptive tree shaping under a fixed token budget. We introduce AdaptiveSpec, a training-free per-step speculative decoding method that adapts both decisions from internal signals already produced during decoding. A per-step margin rule promotes a mismatched draft-proposed token when the ratio of the target's probability on the drafted token to its top-1 probability exceeds a threshold with no dependence on draft length or underlying drafter architecture. A per-step tree policy adjusts the draft tree's depth, width, and node count directly from a fused signal of draft top-1 confidence and a rolling acceptance history capturing recent draft-target agreement, allowing the total draft count to vary rather than only be redistributed. The two adaptations operate on orthogonal axes and compound in effect. Implemented on the SGLang production-grade serving engine, AdaptiveSpec improves throughput over the state-of-the-art autoregressive speculative decoding method EAGLE-3 by up to 56

Background: Air pollution remains a major environmental determinant of health, yet the short-term effects of fine particulate matter (PM2.5) on early respiratory symptoms are poorly characterised because population-scale symptom data are rarely available. We aimed to quantify the association between short-term PM2.5 exposure and nocturnal cough frequency using large-scale mobile sensing data across multiple cities. Methods: We analysed nocturnal audio recordings collected from a widely used mobile sleep application (Sleep Cycle), applying an AI-based cough detection model to estimate nightly cough frequency. Data were aggregated at the city-day level across 32 cities in 12 countries over 500 consecutive days. We used city-specific generalised additive models, multi-city panel regressions adjusting for meteorological conditions and influenza activity, and an event-based analysis during a wildfire episode to assess the association between daily PM2.5 concentrations and nocturnal cough frequency. Findings: Higher daily PM2.5 concentrations were consistently associated with increased nocturnal cough frequency across analytical approaches (relative risk = 1.012, 95% CI 1.006–1.018 per 10 μg/m3). The exposure–response relationship was supralinear, with detectable increases in cough frequency observed even at relatively low PM2.5 levels. The association was strongest at short lags and remained robust in sensitivity analyses. Interpretation: Short-term exposure to PM2.5 is associated with increased nocturnal cough frequency at the population level, suggesting that respiratory symptoms may respond to air pollution at lower exposure levels than previously recognized. Large-scale mobile sensing of respiratory symptoms offers a novel digital biomarker for real-time environmental health surveillance and could complement traditional health monitoring systems.

Ear-worn wearables (aka: earbuds, hearables, or earables) are commonly used by runners for entertainment, and many modern devices also include inertial sensors for user interaction. We propose harnessing the technology embedded in earbuds to capture fundamental aspects of running mechanics and make them available to the wider community of users, outside a lab setting. While other wearables such as insoles or ankle-/sacrum-mounted inertial measurement units have already been presented, ear-worn devices may have a better potential for adoption and therefore offer an optimal compromise between validity of running gait analysis and usability. Thirty healthy participants (18 males, 12 females) ran on an instrumented treadmill (54,000 gait cycles) and floor-mounted force plates (2800 gait cycles) at a variety of speeds. Building on the information brought about by the vibrations transmitted to, and motion of, the head, we devised a gait event detection algorithm and a regression model to predict vertical ground reaction force waveforms. The validation of outcomes against quantities from force plates shows an average mean absolute percentage error of 4.8% on temporal metrics and 9.0% on scalar ground reaction force derived metrics. Additionally, the model tracks the full vertical ground reaction force curve well, achieving an normalized root mean square error of 11.1% on unseen participants. Overall, we show comparable accuracy from an ear-worn consumer device in temporal and kinetic gait parameter estimation to specialist devices, paving the way for accessible running gait monitoring.

Introduction Hypertension is a global health challenge accounting for 8.5 million deaths worldwide despite the availability of low-cost pharmaceutical treatment. About 14.9% of people (9.9 million) registered with primary care practices in England and Wales are prescribed medication for hypertension. However, many patients with hypertension and associated conditions do not take their medications as prescribed. Non-adherence to antihypertensive medication is associated with increased risk of suboptimal blood pressure (BP) control, complications and all-cause mortality, and increased healthcare costs. The Programme on Adherence to Medication (PAM) trial will estimate the effectiveness and cost-effectiveness of a medication adherence intervention in patients prescribed medication for hypertension with poorly controlled blood pressure in primary care. Methods A two-arm multicentre individually randomised controlled parallel group superiority trial recruiting patients prescribed medication for hypertension with poorly controlled BP in primary care practices in England and Wales. The target sample size is 542. Participants in the Intervention group will receive a very brief intervention delivered remotely (by telephone or video call) by a practice nurse or healthcare assistant followed by a digital intervention (text messaging or smartphone app) in addition to usual care; control group participants will receive usual care alone. The primary outcome is systolic BP measured at 12 months. Medication adherence will be measured by chemical adherence testing of urine samples and self-report. An economic evaluation and a process evaluation will be undertaken. Ethics and dissemination The Cambridge East Independent Research Ethics Committee (REC reference 19/EE/0354), the Health Research Authority (HRA) and Health and Care Research Wales (HCRW) approved the trial. The findings will be disseminated to the scientific community, participating practitioners and patients, relevant patient groups and the public using a range of methods, including journal articles, conference presentations, newsletters and the Programme website. Trial registration number The UK’s Clinical Trial Registry ISRCTN82013652

Abstract Stroke volume, the volume of blood ejected by the left ventricle during a contraction, is a key metric of cardiovascular health. Currently, stroke volume is measured in clinic with specialised equipment. While purpose-made wearables exist to measure stroke volume, no solution relies solely on commodity devices. We present a deep learning system for stroke volume estimation from in-ear audio of earbuds. We combine generative self-supervised/transfer learning, a transformer-based autoencoder, to predict average stroke volume in unseen subjects. With data from 23 healthy participants, we compare our estimations to clinically validated device estimations. We achieve a mean absolute error of 5.24 ml, a Pearson correlation of r=0.94 between average predicted stroke volume and average true stroke volume, and a Percentage Error in the limits of agreement of 11.05% (within clinical range for stroke volume measurement devices). These findings open the doors to longitudinal, scalable and affordable cardiovascular measurement out of clinic.

As conversational multimodal AI tools are increasingly adopted to process patient data for health assessment, robust benchmarks are needed to measure progress and expose failure modes under realistic conditions. Despite the importance of respiratory audio for mobile health screening, respiratory audio question answering remains underexplored, with existing studies evaluated narrowly and lacking real-world heterogeneity across modalities, devices, and question types. We hence introduce the Respiratory-Audio Question-Answering (RA-QA) benchmark, including a standardized data generation pipeline, a comprehensive multimodal QA collection, and a unified evaluation protocol. RA-QA harmonizes public RA datasets into a collection of 9 million format-diverse QA pairs covering diagnostic and contextual attributes. We benchmark classical ML baselines alongside multimodal audio-language models, establishing reproducible reference points and showing how current approaches fail under heterogeneity.

Audio-based sleep apnea detection methods hold great potential to improve access to diagnosis, by providing unattended sleep apnea screening at home via sound collected from mobile sensors during sleep. Our research involved a thorough comparison and evaluation of tracheal and ambient microphone recordings for sleep apnea detection with different granularities. Utilising a variety of acoustic representations and sophisticated deep learning architectures, we performed an extensive analysis on the open PSG-Audio dataset, which encompasses over 850 hours of audio data from 194 subjects. For sleep apnea classification, the most effective model showed a 90.8 % accuracy in detecting sleep apnea, 83.3 % accuracy when hypopneic and apneic events were detected separately, and 75.7 % accuracy when apneic events were further divided into three sub-categories. On overnight recordings, the model achieved a sensitivity of 0.93 and a specificity of 1.0 for moderate sleep apnea screening, and a sensitivity of 0.84 and a specificity of 0.97 for severe sleep apnea screening. This research also provided a unique study to compare and combine respiratory sounds from two different types of sensors for sleep apnea detection. The high performance of our model provides a promising avenue for enabling remote diagnosis and monitoring of sleep apnea.

Electrophysiological (ExG) signals offer valuable insights into human physiology, yet building foundation models that generalize across everyday tasks remains challenging due to two key limitations: (i) insufficient data diversity, as most ExG recordings are collected in controlled labs with bulky, expensive devices; and (ii) task-specific model designs that require tailored processing (i.e., targeted frequency filters) and architectures, which limit generalization across tasks. To address these challenges, we introduce an approach for scalable, task-agnostic ExG monitoring in the wild. We collected 50 hours of unobtrusive free-living ExG data with an earphone-based hardware prototype to narrow the data diversity gap. At the core of our approach is Physiology-informed Multi-band Tokenization (PiMT), which decomposes ExG signals into 12 physiology-informed tokens, followed by a reconstruction task to learn robust representations. This enables adaptive feature recognition across the full frequency spectrum while capturing task-relevant information. Experiments on our new HumanSense dataset, the first to enable ExG-based analysis across five human senses, together with four public ExG benchmarks, demonstrate that PiMT consistently outperforms state-of-the-art methods across diverse tasks.

Pervasive sensing enables diverse wearable event detection (WED) applications, but deploying machine learning models on resource-constrained microcontrollers (MCUs) poses significant challenges, particularly in ensuring prediction reliability under data shifts or out-of-distribution (OOD) inputs. While Uncertainty quantification methods offer a way to assess this reliability, many are computationally prohibitive for MCUs, and detecting multiple events concurrently further exacerbates resource constraints. Addressing these combined challenges, this paper presents an uncertainty and resource-aware framework designed for reliable and efficient multi-event WED on MCUs, significantly extending our preliminary work. The proposed framework achieves this by integrating Evidential Deep Learning (EDL) for efficient, single-pass uncertainty estimation with a novel cascade learning architecture. This architecture promotes resource efficiency via: (i) intra-event sharing using uncertainty-aware early exits within a staged model (shallow, medium, deep), allowing simpler samples to terminate inference earlier; and (ii) inter-event sharing using a multi-head design where multiple event detectors share a common backbone, minimizing overhead. System efficiency is further enhanced through MCU-specific optimizations, including targeted architecture search, quantization, efficient uncertainty operator implementation using standard TensorFlow Lite Micro (TFLM) operations, and library footprint reduction. We conducted extensive experiments on four distinct wearable datasets (Oesense, KWS, ECG5000, and HHAR) and two MCU platforms (STM32F446ZE, STM32H747XI), comparing the proposed framework against strong baselines including Deep Ensembles and Vanilla EDL. Results demonstrate the proposed framework's effectiveness, achieving competitive accuracy and uncertainty performance (e.g., up to 22% lower NLL than data augmentation) while drastically reducing resource consumption, offering up to 8.64 & times; faster inference, up to 8.57 & times; lower energy use, and 55% smaller memory footprint compared to ensemble methods. The proposed framework enables the deployment of reliable, uncertainty-aware multi-event detection on a wider range of low-power MCUs.

Language models are remarkably capable at medical question answering, in some cases surpassing the accuracy of general physicians. However, answering questions about wearable health data remains challenging and understudied, as these ubiquitous sensors produce continuous, high-dimensional, and longitudinal data, which is non-trivial to align with text-centric distributions in LLM pretraining. The diversity of sensor modalities and user intents cannot be effectively handled by a fixed reasoning workflow or a single pretrained foundation model. To address these challenges, we propose WEQA, a query-adaptive agent framework that unifies LLM reasoning with specialized wearable analytical and modeling tools. An LLM controller is employed to synthesize execution plans and dynamically route each query to the appropriate combination of sensor analysis and pretrained models, and perform grounded response auditing with external knowledge. We also curate a benchmark spanning four open wearable datasets comprising analytic and predictive tasks in three different health domains. Experiments show that our framework is 24

Test-time adaptation (TTA) offers a compelling remedy for machine learning (ML) models that degrade under domain shifts, improving generalisation with only unlabelled samples. This flexibility suits real deployments, yet conventional evaluations unrealistically assume unbounded processing time, overlooking the accuracy-latency trade-off. As ML increasingly underpins latency-sensitive and user-facing use-cases, temporal pressure constrains the viability of adaptable inference; predictions arriving too late to act on are futile. We introduce , a framework for evaluating TTA under this pressure. It consists of temporal scenarios that model deployment constraints, evaluation protocols that operationalise measurement, and time-contingent utility metrics that quantify the accuracy-latency trade-off. We instantiate the framework with three such metrics: (1) utility for asynchronous streams with hard deadlines, (2) utility for interactive settings where value decays with latency, and (3) utility for budget-constrained deployments. Applying Tempora to seven TTA methods on ImageNet-C across 240 temporal evaluations reveals : conventional rankings do not predict rankings under temporal pressure; ETA, a state-of-the-art method in the conventional setting, falls short in 41.2% of evaluations. The highest-utility method varies with corruption type and temporal pressure, with no clear winner. By enabling systematic evaluation across diverse temporal constraints for the first time, Tempora reveals when and why rankings invert, offering practitioners a lens for method selection and researchers a target for deployable adaptation.

Earable devices, wearables positioned in or around the ear, are undergoing a rapid transformation from audio-centric accessories into multifunctional systems for interaction, contextual awareness, and health monitoring. This evolution is driven by commercial trends emphasizing sensor integration and by a surge of academic interest exploring novel sensing capabilities. Building on the foundation established by earlier surveys, this work presents a timely and comprehensive review of earable research published since 2022. We analyze over one hundred recent studies to characterize this shifting research landscape, identify emerging applications and sensing modalities, and assess progress relative to prior efforts. In doing so, we address three core questions: how has earable research evolved in recent years, what enabling resources are now available, and what opportunities remain for future exploration. Through this survey, we aim to provide both a retrospective and forward-looking view of earable technology as a rapidly expanding frontier in ubiquitous computing. In particular, this review reveals that over the past three years, researchers have discovered a variety of novel sensing principles, developed many new earable sensing applications, enhanced the accuracy of existing sensing tasks, and created substantial new resources to advance research in the field. Based on this, we further discuss open challenges and propose future directions for the next phase of earable research.

Chewing side preference (CSP) has been identified both as a risk factor for temporomandibular disorders (TMD) and behavioral manifestation. Despite TMDs affecting roughly one third of the global population, assessment mainly relies on clinical examinations and self-reports, offering limited insight into everyday jaw function. Continuous CSP monitoring could provide an objective proxy for functional asymmetries. Prior wearable approaches, however, mostly use specialized form factors and demonstrate limited performance. We therefore present CHOMP, the first system for chewing side detection using earphones. Employing OpenEarable 2.0, we collected data from 20 participants with microphones, a bone-conduction microphone, IMU, PPG, and a pressure sensor across eleven foods, five non-chewing activities, and three noise conditions. We apply the Continuous Wavelet Transform to each sensing modality and use the resulting multi-channel scalograms as inputs to CNN-based classifiers. Microphones achieve the strongest single-sensor unit performance, with median F1 scores of 94.5

Data samples used for training often differ from those encountered during fine-tuning and deployment, and while ML models show promise, their performance remains limited when only small annotated datasets are available. Performance often degrades under distribution shifts caused by diverse sensors, populations, and application settings. Although pre-training helps, models frequently encounter out-of-distribution (OOD) data in real-world settings, leading to reduced robustness. Existing adaptation methods usually assume fixed distribution shifts and struggle when multiple types or severities occur. In particular, they overlook shift severity, for example treating adaptation to a large familiar dataset the same as adaptation to a small dataset with a new task, which limits generalisation. To address this, we propose ADAPTOOD, a novel framework that leverages data uncertainty to quantify distribution shift severity and guide fine-tuning for time series. This uncertainty measures how strongly samples from the target deployment distribution deviate from the pre-training distribution, providing a direct signal of OOD severity. Our framework combines this uncertainty with low-rank model updates and adaptive hyperparameter optimisation to improve adaptation. We show that ADAPTOOD achieves up to 7