This study used the TIHM dataset [19], and the participants were recruited from Old Age Psychiatry services and were eligible if they were aged over 50 years, had a verified diagnosis of dementia or mild cognitive impairment, and had the capacity to provide informed consent. Each participant was also required to have a study partner or caregiver who had known them for at least 6 months and could attend research assessments. Individuals with unstable mental health symptoms or terminal illness were excluded. This dataset comprises multimodal sensor data from 56 people living with dementia (aged 70-100 years) monitored over an average of 50 days per participant (2803 total person-days). Of these, 13 participants were included in the analysis because they had valid sliding windows, as described in the Data Analysis for Delirium-Related Anomaly Detection section. Continuous data were collected via passive infrared (PIR) motion sensors, door contact sensors, thermometer, and undermattress sleep mats, providing high temporal resolution for activity patterns, body/skin temperature, and minute-by-minute sleep monitoring (sleep states, respiratory rate, and snoring state). The PIR sensors register ambient motion within a room and do not attribute movement to a specific occupant; recorded activity may therefore include caregiver as well as participant movement. Daily physiological measurements included blood pressure, heart rate, weight, and total body water collected through Bluetooth-enabled devices operated manually by participants or caregivers. The dataset includes validated labels for 6 health-related events (agitation, abnormal blood pressure, abnormal temperature, dehydration, abnormal heart rate, and weight changes) confirmed by their monitoring team. This dataset has previously been used to develop models for detecting agitation and UTIs in dementia populations [12,13].

Data were segmented into sliding windows comprising 10 baseline days and an 11th test day. This sliding window length was selected to construct each participant’s individual behavioral baseline, consistent with prior work showing that 7- to 14-day windows are effective for modelling recent behavioral patterns [25]. Across the 13 participants, the median proportion of missing calendar days within each individual’s recording period was 2.6% (range 0.0%-31.6%). Missing data were handled using a 2-step procedure. A window was deemed invalid and excluded if the test day contained missing data or if at least 5 baseline days (50% of the baseline window) were missing. Overall, a median of 20.5% of candidate windows were discarded (range 3.4%-97.1%). Second, for the remaining valid windows, missing values for each predictor were imputed using the within-window median for that predictor for the same individual. The 2-step rule and the 50% threshold were adopted from previous research on handling missing data in digital health technologies [26,27].

The Isolation Forest and Long Short-Term Memory (LSTM) models were selected for delirium-related anomaly detection because both are suitable for multivariate time-series data. Anomalies detected by either algorithm were treated as proxies for delirium-related events. All markers derived above were used as features for delirium-related anomaly detection within each individual, including daily agitation count, daily sleep quality scores, daily activity entropy rate, daily early warning scores, and daily UTI occurrence. Both models used identical feature sets and preprocessing procedures. The primary analysis used the Isolation Forest algorithm implemented using the scikit-learn library in Python [28], with 20 estimators, a sample size of 100% of the total sample, and a contamination rate of 10%. For each window, the model was trained on the preceding 10 days of data and applied to the subsequent day, which was classified as anomalous or nonanomalous. This rolling-window approach allowed the Isolation Forest to capture recent data distributions, yielding time-sensitive and adaptive predictions relative to each individual’s baseline.

To quantify the contribution of individual features to anomaly detection, a permutation-based importance analysis was conducted for anomaly events. Within each sliding window, values of a single feature were permuted 20 times while holding all other features constant. The importance of that feature was defined as the difference between the average anomaly score of the 11th day obtained from the permuted data and the corresponding score from the original data. This procedure was repeated for each feature within an anomaly event and subsequently across all anomaly events for a given individual. Larger permutation importance scores indicate a greater contribution of the corresponding feature to the detected anomaly.

As a cross-methodological comparison, we additionally conducted an analysis using an LSTM model [29]. The features were min-max scaled to the (0, 1) within individuals for the LSTM. The LSTM was implemented as a sequence-to-sequence autoencoder comprising an encoder LSTM layer (32 units) whose final hidden and cell states were passed through a RepeatVector of length 10 to a decoder LSTM layer (32 units, return sequences enabled), followed by a TimeDistributed dense output layer with 5 units to reconstruct the input sequence. One model was trained per participant on that participant’s baseline windows for 20 epochs with a batch size of 4, using the Adam optimizer (default learning rate of 1e-3, no scheduler) and mean squared error as the reconstruction objective. Specifically, for each participant, one LSTM encoder-decoder model was trained on all available baseline windows pooled together, learning to reconstruct normal 10-day sequences for that individual. At scoring time, the trained individualized model reconstructed both the 10-day baseline sequence and the test-day data within each sliding window, yielding a baseline and a test reconstruction error, respectively. The anomaly score was defined as the difference between these 2 errors. Thus, each test day was evaluated relative to its own preceding 10-day baseline, preserving window-level temporal context at inference while being more computationally efficient than training a separate model for each window. The anomaly quantile for the LSTM was set at 15% to match that of the Isolation Forest. The agreement rate between the two methods was calculated. The data analysis workflow is presented in Figure 2.

The TIHM study received ethics approval from the London-Surrey Borders Research Ethics Committee (TIHM 1.5 REC: 19/LO/0102), and the dataset has been made public with participants fully anonymized. Participants provided informed consent prior to participation. Each participant was required to have a study partner or caregiver who had known them for at least 6 months and was able to attend research assessments. Where participants were unable to reliably report information regarding their health, the study partner or caregiver completed the relevant assessments on their behalf. All the participants have granted the publication of this dataset. More details can be retrieved from the original paper [19].

This study demonstrates the feasibility of using theory-driven markers to detect delirium-related patterns in home settings for individuals with dementia, addressing the challenges of clinical delirium assessment, particularly in dementia populations. By applying unsupervised anomaly detection algorithms to within-individual data using sliding windows, this approach ensures the time sensitivity of detection while accounting for individual differences in symptom deviations. Although validated labels for delirium episodes were not available, the detection of multivariate anomalies in home sensor data using markers aligned with CAM criteria provides confidence that these anomalies correspond to known delirium symptoms. Furthermore, several additional findings discussed below further support the validity of this detection approach.

First, we observed that numerous detected delirium-related episodes show patterns of temporal clustering, which align with known delirium patterns [31]. Second, we observed that for some clustered delirium-related episodes detected, the composing markers demonstrate concurrent deteriorations. This also aligns with previous findings that the mean number of CAM symptoms preceding delirium episodes is 2.4, and the mean number is 3 after episodes [32]. Our result patterns match patterns normally observed in delirium episodes in older adult residents, which provides some evidence for the effectiveness of detection.

We observed variation in agreement rates across individuals between the two anomaly detection algorithms, which reflects methodological differences rather than limitations in the anomaly detection process itself. LSTM models, with their focus on sequential data, are designed to capture longitudinal changes and temporal dependencies in the data [33]. As a result, if the previous data shows more fluctuation, the LSTM model may be more tolerant of deviations, since it inherently learns the patterns of variability over time. In contrast, the Isolation Forest algorithm focuses on the global distribution and deviations of markers within fixed time windows, without considering temporal sequence [34]. As a result, it is more sensitive to extreme deviations from the overall baseline. Given that this study is a proof of concept with flexible modular algorithms for anomaly detection, future datasets with validated labels will be necessary to determine which algorithm is best suited for delirium-related anomaly detection.

However, there are important limitations in this study. First and most importantly, the lack of ground truth labels makes it impossible to determine the accuracy and other performance measures such as sensitivity and specificity of the detection algorithms. While this study provides compelling evidence that the detected delirium-related patterns align with existing literature and that our markers are consistent with CAM criteria, it goes beyond a proof of concept. A key contribution of this study is the introduction of a novel, data-driven approach to delirium-related anomaly detection in home settings, highlighting the potential for continuous, individualized monitoring of delirium-related patterns using theory-driven markers. Second, while we aimed to construct markers that closely align with CAM items, future efforts should focus on collecting more comprehensive data to enable the derivation of additional relevant markers that are unavailable in the current dataset (eg, inattention and altered level of consciousness) and ensure that the proxy markers accurately measure the same constructs as those defined by the CAM items. For instance, circadian disorientation was also included in the disorientation marker [9,35], but this is unavailable in the current dataset. There is a caveat that behavioral proxies such as agitation and entropy cannot fully capture the complexity of cognitive symptoms and therefore do not constitute direct substitutes for clinical cognitive assessment. Rather than replacing individual CAM criteria, our aim is to derive an objective, continuous, and nonintrusive approximation using smart home data, with the goal of achieving detection performance comparable to the CAM.

Apart from the lack of validated labels, the current dataset has several additional limitations that should be addressed in future data collection. First, sleep data were available for only a subset of participants because the mattress sensor was introduced later in the study, and some participants did not consent to its installation. Future translational work should investigate these barriers more directly, particularly whether concerns relate to comfort, privacy, or other acceptability reasons. A further limitation is that the duration of sensor recording varied across participants, resulting in an unequal number of valid baseline–test window pairs per individual. Participants with shorter recording periods contributed fewer windows, which may have reduced both the opportunity to detect anomalies and the stability of the personalized behavioral baseline. Although the 10-day sliding window approach does not in principle require very long longitudinal data, longer monitoring periods would allow a more representative characterization of typical behavior and make the anomaly thresholds more meaningful. These differences in recording duration partly arose from the rolling recruitment design of the ongoing study, which could be improved in future study design, and partly from participant dropout or death, which is an inherent challenge in longitudinal research involving older adults living with dementia.

In addition, several factors may influence how sensor-derived activity patterns are interpreted. The PIR sensors capture motion as discrete events and do not attribute movement to a specific individual, so caregiver presence may contribute to recorded activity. This concern is mitigated in 2 ways. First, because anomalies are defined as deviations from each participant’s own preceding baseline, habitual caregiver activity is incorporated into that baseline and does not by itself generate anomalies. Second, of the 5 markers, only activity entropy is derived purely from ambient motion; sleep quality, early warning score, UTI, and agitation are substantially patient specific, and our feature importance analysis (Figure 4) indicates that detection is driven by a combination of these markers rather than by entropy alone. Nevertheless, because the open TIHM dataset does not record living arrangements or the proportion of time a participant is unaccompanied, we could not stratify analyses by co-residence or occupancy, and contextual influences cannot be fully ruled out. We recommend that future data collection enable patient-specific measurement of activity, for example, through patient-specific PIR sensing, to ensure that recorded signals correspond only to the target patient.

Nevertheless, this study contributes to detecting delirium-related anomalies in home settings in several important aspects. First, this is the first study to work on detecting delirium-related anomalies in smart home settings and provides evidence of viability, which has huge potential for prevention and early intervention for people living with dementia and reducing costs for hospital admissions. Second, our individualized, unsupervised detection framework is well aligned with real-world implementation contexts, in which delirium-related episodes are difficult to clinically validate in a timely manner. By modelling within-person baselines rather than relying on global population models, this approach accommodates substantial interindividual heterogeneity and captures dynamic shifts in an individual’s baseline as biological states and daily behaviors evolve over time. This study lays a strong foundation for future work, including developing data collection protocols that better align with the goal of delirium-related anomaly detection, and algorithm improvement and validation through clinical partnerships.

This proof-of-concept study demonstrates the feasibility and promise of using smart home data to detect anomalies consistent with delirium superimposed on dementia, with potential value in reducing health care costs through early intervention. The limitations of lacking clinically confirmed labels and the strengths of detecting anomalies while considering individual differences are acknowledged. Future work requires data collection with a focus on delirium prevention, including data that allow features better aligned with delirium symptoms, as well as clinically confirmed labels for ground truth validation.