Smart home Internet of Things-based behavioural analysis for early detection of cognitive decline: toward Saudi future vision

Introduction

Today, the early identification of health-related risks is becoming a significant focus for medical science and researchers. Medical science, in coordination with artificial intelligence (AI) and related fields including machine learning (ML), deep learning (DL), natural language processing (NLP), and neural networks, has produced data-driven approaches for analysing complex health-related patterns. Cognitive health is one such domain that requires urgent attention from researchers and medical experts due to its gradual and often subtle decline. After analysing several studies, it becomes evident that the outcomes of such research largely depend on the nature of the datasets and behavioural factors under consideration. Accordingly, this study focuses on observing and analysing human behavioural activities using IoT-based smart-home environments.

Unobtrusive and constant monitoring of human behaviour has changed as a result of the Internet of Things’ (IoT) quick combination into modern smart houses. Dementia and cognitive impairment are becoming major public health issues due to the world’s aging population growing at an unparalleled rate (1,2). Early recognition of minor behavioural deviations may support timely intervention and risk awareness, helping to mitigate severe cognitive decline. Conventional diagnostic methods, such as neuropsychological testing and clinical evaluations, are time-consuming, costly, and often reliant on subjective judgement. In contrast, IoT-based behavioural monitoring offers a non-invasive and data-driven means of analysing daily functional variations over time (3). Smart-home IoT sensors—such as motion sensors, pressure sensors, and appliance sensors—continuously track residents’ activity patterns and provide rich behavioural data streams. This analysis enables the identification of behavioural irregularities, including sleep disturbances, reduced mobility, or altered daily routines, which have been reported in association with mild cognitive impairment (MCI) and Alzheimer’s disease (AD) cohorts (4). However, transforming high-dimensional and dynamic temporal sensor data into meaningful behavioural indicators remains a key challenge, as illustrated in Figure 1.

Figure 1 Daily smart home activities. HVAC, heating, ventilation, and air conditioning.

IoT-based behavioural monitoring and cognitive health

Recent years have seen a rise in research interest in the behavioural monitoring of cognitive decline, particularly with the introduction of smart-home and IoT-based technologies. Rather than focusing on clinical diagnosis, many recent studies investigate how daily activity patterns can serve as behavioural indicators associated with cognitive health and risk (5,6). According to multiple research analyses, information derived from activities of daily living (ADLs) can provide valuable insights into long-term cognitive trends (7,8). Self-reports and infrequent medical examinations remain the mainstays of conventional approaches; however, these methods often overlook subtle, progressive behavioural deviations that may emerge over extended periods (9).

Researchers at the University of Washington’s CASAS Smart Home Project were among the first to employ ambient sensors for tracking daily activities in older adults (4). Their findings demonstrated that routine behavioural metrics—such as sleep duration, kitchen usage, and mobility patterns—can be associated with cognitive impairment risk categories rather than serving as definitive diagnostic outcomes (10). Similarly, MIT’s AgeLab and the Oregon Center for Aging and Technology (ORCATECH) deployed long-term sensor networks to observe daily routines and behavioural changes, establishing the feasibility of continuous, in-home behavioural monitoring in smart environments (11,12) .

Beyond academic prototypes, businesses and healthcare organisations have begun incorporating IoT-based behavioural analytics for health-related monitoring (5,13). Research involving wearable sensors, voice assistants, and smart devices shows that multimodal data fusion can enhance the robustness of behavioural pattern analysis, for example by combining motion sensor data with appliance usage to capture routine consistency, forgetfulness, or task interruptions. ML and DL methods are increasingly applied for this purpose, including Hidden Markov Models (HMM), support vector machines (SVMs), and long short-term memory (LSTM) networks for modelling temporal activity data and identifying behavioural anomalies (14,15). These approaches aim to flag deviations from baseline behaviour rather than perform direct clinical diagnosis. However, existing solutions continue to face challenges related to scalability, personalisation, and data privacy (6).

In parallel with North American and European initiatives, several Asian and regional research efforts have explored smart-home and ambient-assisted living technologies for ageing populations (16,17). Studies from East and Southeast Asia have investigated passive sensing, ADL monitoring, and privacy-preserving IoT deployments in eldercare settings, highlighting region-specific challenges related to cultural routines, housing layouts, and long-term technology adoption (12,18). These global efforts collectively demonstrate the growing international interest in behavioural monitoring frameworks for cognitive health research.

In addition, much of the prior work has focused on individual sensor modalities or relatively small datasets, which limits generalisability across diverse living environments. Many studies do not sufficiently address the dynamic and evolving nature of human behaviour over time, nor the integration of adaptive systems that personalise models to individual users (19,20). Ethical and data governance challenges—such as informed consent, data ownership, and algorithmic transparency—have also received limited attention in earlier literature. The HEALNET (Home Environment Assisted Learning Network) framework builds upon these prior efforts by integrating multiple sensor streams into a unified behavioural analysis pipeline, promoting continuous learning, user-specific behaviour modelling, and privacy-aware data management. This positioning emphasises behavioural risk screening and research-oriented assessment rather than clinical decision-making, enabling a more scalable and ethically grounded approach to smart-home cognitive monitoring.

ML-enabled approaches

ML approaches have been widely used to detect patterns of cognitive decline by evaluating behavioural and sensory data (7,21). SVMs, Random Forests (RFs), and logistic regression (LR) are popular techniques in artificial intelligence (AI) and related subfields for classifying activity patterns, identifying anomalies, and differentiating between people with and without disabilities. ML-based models are best suited for feature extraction, pattern recognition, and decision-making from structured data collected from smart home environments. However, these methods typically rely on manually engineered features and labelled datasets, which can limit their adaptability and scalability across heterogeneous households and long-term deployments.

DL-based approaches

In addition, we provide a DL-based ML approach here. Therefore, DL models, specifically convolutional neural networks (CNNs) and LSTM networks, have revolutionized the monitoring of cognitive health by automatically learning hierarchical representations from complex sensor data (14,19). While LSTMs efficiently represent temporal patterns in daily behaviour, CNNs identify spatial correlations in activity data. Compared with traditional ML techniques, DL-based approaches are better suited for identifying subtle and gradual behavioural changes associated with cognitive risk. Nevertheless, DL models often require substantial computational resources, large datasets, and careful hyperparameter tuning to avoid overfitting in real-world smart-home settings.

Ensemble and explainable AI (XAI)-based approaches

Recently, ensemble learning and XAI paradigms have arisen that utilize the predictive capabilities of ensembles while considering interpretation for clinical use (22,23). Ensemble techniques such as Ada Boost, Bagging, and Stacked Simplification enhance performance and robustness by combining heterogeneous apprentices. In the context of behavioural cognitive monitoring, XAI contributes to ethical and trustworthy AI deployment by reducing the risk of opaque decision-making, particularly when analysing sensitive behavioural data. These approaches help bridge the gap between performance and reliability, making AI-based behavioural monitoring frameworks more suitable for real-world smart-home research environments, as summarised in Table 1.

Research gaps and motivation

Despite significant progress, several research gaps remain:

• Existing limited integration of real-time behavioural data with cognitive assessment models.
• Inadequate use of DL (CNN, LSTM) for long-term activity tracking.
• Lack of personalized cognitive decline prediction models.
• Minimal research on privacy-preserving IoT frameworks.
• Insufficient multimodal sensor fusion for accurate behavioural analysis.

These limitations motivate the need for an integrated, adaptive, and ethically grounded framework capable of continuous behavioural monitoring in real-world smart-home environments.

Motivation, contributions, and ethical implications of this research

Mental disorders such as dementia and AD are often identified at relatively advanced stages, after substantial neurodegeneration has already occurred. This reality highlights the pressing need for continuous, non-invasive monitoring approaches capable of detecting early behavioural deviations associated with cognitive decline. Smart-home IoT environments enable long-term observation of daily activities and help bridge the gap between episodic clinical observation and real-world behaviour. Motivated by this need, the present study focuses on behavioural risk screening rather than diagnostic decision-making, aiming to support early-stage research and monitoring.

Contribution of this study:

By developing a framework that analyses behavioural data collected through smart-home IoT sensors, this research contributes to the expanding field of digital healthcare in the following ways:

• It proposes a task-driven behavioural analysis framework for modelling daily activity patterns associated with cognitive risk;
• It integrates spatial, temporal, and statistical representations using CNN, LSTM, and ensemble learning techniques;
• It provides an experimentally validated, research-stage framework for behavioural cognitive-risk screening using publicly available datasets.

Ethical considerations are central to the HEALNET design. Data privacy, security, and informed consent are prioritised through anonymisation, secure handling, and ethically sourced datasets. The framework adheres to transparency and minimal intrusiveness while considering algorithmic fairness and bias to support responsible AI deployment.

The proposed system—HEALNET—provides the task-specific integration of heterogeneous behavioural representations for longitudinal cognitive-risk screening using passive IoT sensing. Its modular design supports scalability, cost-effectiveness, and integration with existing smart-home infrastructures, particularly within Saudi smart-home contexts.

This paper is logically organized into five sections. Section “Introduction” presents the introduction, background, motivations, and objectives of this research. Section “Methods” describes the research methodology, datasets, and algorithmic approach, including the proposed HEALNET model. Section “Results” describes the experimental design and results. Section “Discussion” provides the discussions and evaluation of the model with the limitations of the study, while Section “Conclusions” summarizes the main findings and discusses future directions of the study and potential future developments.

Methods

This study uses an experimental and data-driven research methodology to investigate the routine activities or behavioural patterns of smart home IoT devices in order to facilitate the early identification of cognitive impairment. The suggested framework, HEALNET, integrates many sensor data streams, including motion sensors, temperature sensors, wearable activity trackers, and gadget usage antiquities, into a unified data dispensation structure. By continuously capturing people’s daily activities indoors, these sensors generate time-stamped behavioural data that describes routine consistency, interaction patterns, and crusade patterns. Pre-processing for data cleaning includes noise reduction, normalization, and subdivision to remove missing data. Following pre-processing, the data is subjected to feature extraction based on variables such as the length of the activity, the frequency of room changes, the regularity of sleep, and the patterns of the device interface.

These behavioural traits are fed into ML models such as SVMs, RFs, and LSTM neural networks, as seen in Figure 2. A number of models are available; these models are chosen based on their ability to capture temporal dependence and identify subtle behavioural abnormalities. The image illustrates the HEALNET workflow, which shows the continuous process from collecting data from smart homes to detecting cognitive decline. Sensor data collection (motion, temperature, wearables) is followed by data preparation (noise reduction and normalization). After that, it displays feature extraction and fusion, where CNN, LSTM, RF, and SVM modules are used to assess behavioural patterns. Finally, the GIF illustrates how HEALNET continuously familiarizes itself while upholding data privacy and ethical compliance by displaying adaptive feedback and prediction as illustrated in the Figure 1.

Figure 2 Outlines of the research methodology used in study. HEALNET, Home Environment Assisted Learning Network; LSTM, long short-term memory; SVM, support vector machine.

Cross-validation type performance parameters are being used for training the models utilizing marked datasets, such the Centre for Advanced Studies in Adaptive Systems (CASAS) smart home dataset, to quantify classification accuracy, precision, and recall. This study documents baselines and behavioural changes over time using both supervised and unsupervised learning techniques. Lastly, the framework can adapt to a person’s evolving lifestyle by using feedback to continuously update the model. Because HEALNET offers ethical measures through data collection, encryption techniques, and user informed authorization, it is a safe and responsible way to operate in a true smart home context.

Classification algorithm

The HEALNET framework’s analytical core, a decision-making algorithm, is used in this study to precisely identify patterns of cognitive deficiency from rich behavioural data. SVMs and LSTM neural networks are combined in the system’s hybrid ML model to achieve the ideal balance between interpretation and temporal learning capacity. Here, SVMs are employed for inactive behaviour categorization, effectively differentiating between patterns of normal and pathological activity based on spatial characteristics and frequency-based parameters including routine consistency, movement intensity, and frequency of gadget usage. Additionally, in order for the system to detect slow-onset behavioural abnormalities that might point to cognitive impairment over time, the LSTM network also processes sequential data to capture temporal dependencies. By recording both immediate and gradual behavioural changes, the integration of these conventional individual algorithms enhances accuracy and robustness. To enhance generalizability and prevent overfitting, models are trained using k-fold cross-validation on interpreted datasets, like the CASAS smart home dataset. Adaptive learning is used to continuously update the learned classifier’s predictions of potential cognitive states based on fresh data. In the context of an interactive smart home, HEALNET can offer continuous, real-time cognitive monitoring thanks to this two-layer classification technique.

SVMs

In order to categorize data, SVMs locate the hyperplane that optimizes the margin between distinct classes. The basis for the decision function is:

f(x)=wTx+b[1]

where w is the weight vector and b is the bias. SVM tries to minimize:

min12‖w‖2[2]

provided that every data point is accurately categorized with the greatest margin.

RF

An ensemble learning technique called RF builds a large number of decision trees during training and produces the mean prediction for regression tasks or the mode of the classes for classification. Random subsets of characteristics are chosen, and bootstrapped samples are used to construct each tree in the forest. The ultimate forecast y is:

y=1T∑t=1Tft(x)[3]

where T is the number of trees, f_t(x) represents the t-th tree’s prediction, and x represents the input feature vector. Randomization enhances model robustness and helps avoid overfitting in both data and feature selection.

LR

This method uses the logistic function to represent the likelihood of a binary result. The representation in mathematics is:

P(Y=1|X)=11+e−(β0+β1X1+β2X2+⋯+βnXn)[4]

where the coefficients to be estimated are denoted by β.

CNNs

A family of DL models called CNNs is mainly employed for text and picture categorization applications. CNNs use convolutional filters to extract local information from input text sequences in order to identify emotions. The definition of the convolution operation is:

y=f(x∗w+b)[5]

where x is the bias, x is the filter, and * indicates convolution. CNNs can discover patterns in words and phrases to detect emotions in text because they are good at capturing spatial hierarchies in data.

Recurrent Neural Networks (RNNs)

Because RNNs can handle sequential input, they are perfect for jobs where word order counts, such as text emotion recognition. RNNs can handle sequences of any length because they keep track of past inputs in a hidden state. The state that is concealed at time t is provided by:

ht=f(Wh·h+t−1Wx·xt+b)[6]

where the current input is x_t, the prior hidden state is h_t−1, and the weight matrices are W_H and W_x. RNNs may discover context-dependent patterns in text data because to this structure.

LSTM

LSTM networks are a special type of RNN capable of learning long-term dependencies in sequential data. Each LSTM unit maintains a cell state C_t updated through gates:

Ct=ft⊙Ct−1+it⊙Ct˜[7]

where f_t, i_t, and Ct˜ are the forget, input, and candidate gates, respectively. LSTMs are ideal for modelling time-dependent behavioural data, such as movement patterns or routine shifts, allowing HEALNET to detect gradual changes indicative of cognitive decline.

Experiment setup

The experimental setup of this study was designed to validate the operation of the HEALNET framework for detecting early cognitive decline based on the analysis of behavioural patterns. The experiments used the publicly released CASAS smart home dataset created by the University of Washington. This database consists of multimodal sensor data that monitors the daily activities of older adults in real home environments. It includes motion, door, light, and appliance sensors as well as time-stamped activity annotations that are used as reference labels for model inference. Moreover, this now to deal with missing values, eliminate sensor noise, and normalize all numerical features, data pre-processing was employed. To reservation time dependency, the behavioural sequences were detached into daily action windows. A 70% training set, 15% validation set, and 15% test set were created from the pre-processed data. Rules for process time, transition frequency, device usage patterns, and temporal regularity were applied to build feature vectors. Python (TensorFlow and Scikit-learn frameworks) was used to run all classification algorithms (SVM, RF, LR, CNN, RNN, and LSTM) on a powerful machine equipped with an Intel i7 CPU, 32 GB of RAM, and an NVIDIA RTX. To ensure the optimal model configuration, grid search and cross-validation were used to alter the hyperparameters. Performance of the proposed model was assessed using performance parameter metrics including precision and precision.

Software requirements

To speed up the processing of big data sets, the study was carried out on a high-performance computer system using an NVIDIA GPU (such the Tesla V100). This computer’s Intel Xeon CPU and 128 GB of RAM allowable it to professionally complete complicated calculations. The application was mostly written in Python programming language, and the data was processed, analysed, and ML models were constructed using libraries like TensorFlow, Scikit-learn, and Pandas. The Git allowed for version control and teamwork, while the Jupyter Notebook environment allowed for interactive coding and conception.

Statistical analysis

The AI model’s performance was assessed using a number of important performance metrics. Accuracy assessed the overall precision of the predictions, whereas sensitivity (or recall) assessed the model’s capacity to detect true positives. The model’s specificity shows how effectively it could observe real negatives. The receiver operating characteristic (ROC) curve’s area under the curve (AUC) was used to evaluate the model’s cultivated power. To calculate the percentage of real positive predictions among all positive forecasts, accuracy was also taken into account for better results. Each of these metrics offers a thorough evaluation of the model’s beneficial usefulness and forecast precision.

Accuracy

The total accuracy of the model’s predictions is measured by accuracy. It has the following definition:

Accuracy=TP+TNTP+TN+FP+FN[8]

where the values of TP, TN, and FP represent true positives, true negatives, and false negatives, respectively.

Precision

Precision evaluates positive prediction accuracy, which is described as:

Precision=TPTP+FP[9]

In this case, genuine positives are represented by TP, true negatives by TN, false positives by FP, and false negatives by FN. The percentage of accurate predictions made out of all the cases is called accuracy.

Recall

Recall, also known as sensitivity, gauges how well the model can detect real positives:

Recall=TPTP+FN[10]

It displays the model’s accuracy in capturing real-world positive cases.

F1-score

The harmonic mean of recall and accuracy, which balances both, is the F1-Score:

F1-score=2×Precision×RecallPrecision+Recall[11]

It is particularly effective when there is an unequal class distribution.

ROC AUC

AUC, or ROC area under curve, measures how well the model can differentiate between classes:

AUC=∫01TPR(FPR)dFPR[12]

At different thresholds, the true positive rate (TPR) is shown versus the false positive rate (FPR).

Specificity

The percentage of genuine negatives that are accurately detected is measured by specificity.

Specificity=TNTN+FP[13]

It focuses on how well the model can recognize negative instances.

Confusion matrix

The confusion matrix is a powerful performance measurement tool for classification models. It graphically displays the relationship between the correct and predicted labels. The confusion matrix has four main parameters: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). It is used to measure the model’s precision, accuracy, recall, and F1-score in a more detailed manner than precision. In the case of IoT behaviour-based cognitive impairment detection, this metric guarantees a strong separation between healthy, mildly impaired, and cognitively impaired subjects.

The mathematical expression of accuracy from the confusion matrix is as follows:

Accuracy=TP+TNTP+TN+FP+FN[14]

This approach ensures that misclassifications (FP and FN) are minimized, which is important in health predictions.

Correlation matrix

A correlation matrix measures the strength of the relationship between multiple variables and provides evidence of how the attributes interact. Each cell contains a Pearson correlation coefficient (r) that varies from −1 to +1. A strong positive correlation indicates a simultaneous increase in two attributes, while a negative correlation indicates a decrease in one and an increase in the other. In this study, the correlation between sensor activity, movement patterns, sleep duration, and device usage shows how behavioural habits are associated with the risk of cognitive decline.

The correlation coefficient is calculated as follows:

rxy=∑(xi−x¯)(yi−y¯)∑(xi−x¯)2∑(yi−y¯)2[15]

This matrix helps prevent attribute selection, data reduction, and multicollinearity during model training.

Dataset description

CASAS Smart Home Datasets (University of Washington)

The smart home dataset from CASAS is Real-world IoT sensor data from smart home environments (34) basically originated from CASAS, Washington State University.

The CASAS Smart Home dataset is an exhaustive set of real-world sensor data collected from 189 community smart home testbeds that are intended to research human activity and health monitoring. The dataset captures ambient sensor streams like motion detectors (PIR), temperature, door sensors, and appliance sensors located inside actual residential settings. Each home in the dataset contains about 30–80 sensors. It offers timestamped sequences of sensor events with aligned annotated activity labels (e.g., “Cooking”, “Eating”, “Sleeping”) that can be translated to behavioural patterns of ADLs. The collected data spans several weeks to months for each participant resulting in Thousands of sensor events per day over an 18-year span resulting in volume 1.4 TB of sensor data.

Significance of dataset utilization in this study

Here, in this study, the CASAS dataset was used as the main evaluation and training standard dataset for the HEALNET model. CASAS’s multimodal, time-stamped format allowed for the following:

• Spatial feature discovery using CNN (detection of activity distributions between rooms);
• Temporal sequence modelling with LSTM (detecting daily routine coherence and inconsistencies);
• Classification of behaviour using SVM and RF ensemble strategies for accurate cognitive state prediction.

The training pipeline utilized 70% training, 15% validation, and 15% testing splits of CASAS data. Cross-validation helped in generalizing the model and avoiding overfitting. Accuracy, precision, recall, F1-score, and ROC-AUC were calculated as performance measures using labelled data available in CASAS activity annotations.

Ethical considerations

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

CASAS dataset characteristics

The following Table 2 shows the used CASAS dataset characteristics in this study.

Pre-processing and feature extraction

Prior to analysis, the raw event logs were:

• Cleaned—deleting duplicate events, fixing timestamps, and managing missing data;
• Segmented—divided into hourly or daily windows to observe temporal patterns of behaviour;
• Normalized—here all sensor initiations were normalized to eradicate variations in activity frequencies;
• Feature-engineered—derived behavioural attributes including:
  • Activity duration and frequency;
  • Transition probability between rooms;
  • Sleep duration and irregularity metrics;
  • Appliance usage frequency and time-of-day patterns.

These features were used as inputs to the proposed model HEALNET’s CNN-LSTM hybrid model for the recognition of cognitive decline.

Proposed model

This study proposes HEALNET, a DL designed for early signs detection of cognitive decline through behavioural analysis using smart IoT based data. This model combines temporal, spatial, and statistical representations of residents’ daily activities to provide an interpretable and privacy-preserving prediction system. The overall workflow of HEALNET is presented in Figure 3. The dataset denoted consists of multivariate time-series data collected from ambient smart home sensors (motion, door, light, temperature, and appliance sensors).

Figure 3 Proposed architecture of the HEALNET framework for cognitive decline detection using IoT-based behavioural data. CNN, convolutional neural network; DP-SGD, Differentially Private Stochastic Gradient Descent; IoT, Internet of Things; LSTM, long short-term memory; RF, Random Forest; SVM, support vector machine.

Each sample represents a time window of activity sequences, corresponds to the cognitive state: healthy, MCI, or cognitive decline as below:

Let D={(X(i)),y(i)}i=1N be the dataset, where each sample X(i)={xt(i)}t=1Ti is a multivariate time series of sensor readings and y(i)=∈{0,1,2} (0: healthy, 1: MCI, 2: cognitive decline).

Let ∅(·) denote preprocessing/feature-extraction, ⊕ concatenation, and ‖·‖ the Euclidean norm. Scalars α, β, γ are loss weights.

Algorithm: evaluating predictive accuracy and clinical utility of an AI model

Pre-processing & feature extraction

Pre-processing of raw sensor readings includes:

• Filtering noise and timestamp alignment to sync sensor data that isn’t in sync.
• Segmentation with a sliding window (duration =60 s with 50% overlap) to obtain ongoing behavioural context.
• Extraction of features, which involves:
• Frequency of activity transitions
• Intensity of motion (average, variability, and randomness)
• Duration of sleep and periods of inactivity
• Patterns of appliance usage the z-score normalization is applied to these characteristics to eliminate subject bias as follows:

Sensor reading at time t: st∈ℝm.

Windowed feature vector for window w (length L):

fw=ϕ({st}t=(w−1)L+1wL)∈ℝd[16]

Common features: activity counts, transition frequencies, mean/std of motion, sleep-duration, appliance-use frequency.

Normalize:

fw˜=fw−μfσf[17]

where μ_f, σ_f are feature-wise empirical mean and std.

Multimodal feature learning

The pre-processed data are categorized into three complementary streams:

• Spatial stream (CNN encoder): sensor activations are used to create activity mappings in the spatial stream (CNN encoder), where each channel represents a specific sensor location. In home settings, a CNN recognizes spatial relationships by spotting patterns like less movement or
• Fewer room transitions. Temporal stream (LSTM encoder): Sequential event data are analysed using a LSTM network to capture temporal correlations, such as changes in daily routines or irregular sleep-wake cycles.
• Static stream (RF/SVM): RF and SVM models are used to examine gathered statistical parameters (such average movement count and total door occurrences) for accurate and intelligible classification.

The outputs of these three components are combined into spatial, temporal, and static embeddings by a feature fusion layer in the manner described below:

Split features into static f(S) and temporal sequences F(T)=[f1˜,⋯,fW˜].

Apply spatial encoder (CNN) on activity maps:

zCNN=CNN(reshape(F(T)))[18]

Apply temporal encoder (LSTM) on sequence:

ht=LSTM(ft˜,ht−1),zLSTM=hW[19]

Apply static classifier features via RF/SVM:

zRF=RF(f(S)),zSVM=SVM_score(f(S))[20]

Concatenate fused representation:

z=zCNN⊕zLSTM⊕zRF⊕zSVM[21]

Optional attention weighting:

αk=exp(wkTzk)∑jexp(wjTzj),zfused=∑kαkzk[22]

Classification & anomaly detection

To produce class probabilities for each of the three cognitive states, a SoftMax classifier is represented, A Mahalanobis distance score between present behaviour and baseline healthy patterns is also calculated by an anomaly detection layer. An early warning sign for potential cognitive deterioration is triggered by high deviation as below:

Softmax probabilities (for multi-class):

p^=softmax(Wozfused+bo),y^=argmaxp^[23]

Anomaly/decline score (reconstruction or Mahalanobis):

δanomaly(X)=‖zfused−μclass‖C−1(Mahalanobis)[24]

where µ_class and C estimated from training healthy examples.

Loss function (training objective)

The HEALNET framework’s training goal is to guarantee precise, reliable, and privacy-preserving predictions for the classification of cognitive states. A number of components come together to form the overall loss function L_{total}, which minimizes overfitting and noise sensitivity while directing the model toward robust learning as below:

Total loss:

L=αLCE(p^,y)+βLconsistency+γLpriv[25]

Cross-entropy:

LCE=−∑c1{y=c}logpc^[26]

Temporal consistency (smoothness) — penalize abrupt changes in representation:

Lconsistency=1W−1∑t=1W−1‖ht+1−ht‖2[27]

Privacy regularizer (DP-inspired): add noise budget penalty or gradient clipping:

Lpriv=λ∑pclip(‖∇pLCE‖,C)[28]

(Implement via DP-SGD for deployment.)

Regularization (weight decay):

L←L+η‖Θ‖2[29]

Optimization & training

A composite loss function is minimized during training:

• Cross-entropy loss for classification accuracy.
• Loss of temporal smoothness to ensure uniformity across successive forecasts.
• Differential privacy [Differentially Private Stochastic Gradient Descent (DP-SGD)], a formalization of privacy that limits the disclosure of private behavioural data.

The model’s parameters are optimized using an optimizer with adaptive learning rates. For ensemble models (RF and SVM), grid search is used to modify hyperparameters like a tree depth and kernel type. The final predictions are derived using a weighted ensemble fusion of all sub-model outputs as below:

Neural parts: optimize Θ via mini-batch Adam:

Θ←Θ−Adam(∇ΘL)[30]

RF: bootstrap aggregation; optimize tree depth and number via grid search.

SVM: solve dual problem (if used) or use lib SVM with RBF/kernel

K(xi,yi)=exp(−γ‖xi−xj‖2)[31]

Ensemble decision rule

HEALNET uses an ensemble decision method that integrates outputs from several models, including CNN, LSTM, RF, and SVM, to increase predictive resilience and lessen bias from individual classifiers. The ensemble layer uses weighted fusion to integrate the probabilistic estimates of a subject’s cognitive state provided by each model as below:

Let outputs be class probabilities/scores from models:

p^(NN),p^(RF),S(SVM)[32]

Convert SVM score to probability via Platt scaling. Final ensemble probability:

p^ens=wNNp^(NN)+wRFp^(RF)+wSVMp^(SVM)[33]

with w’s normalized (Σw =1). Decision:

y^=argmaxcp^ens,c[34]

Thresholding & alerting

HEALNET’s thresholding and alerting system is intended to convert probabilistic model results into clinically useful judgments. A decision threshold is used to assess if a subject is displaying behavioural indicators of cognitive deterioration once the ensemble layer has generated the final probability distribution as below:

Set threshold τ on anomaly score or class probability for clinical alert:

Alertifδanomaly(X)>τorp^ens,decline>τp[35]

Choose τ based on ROC operating point (maximize sensitivity under acceptable specificity).

Evaluation metrics (for reporting)

The model performance is evaluated using standard metrics: accuracy, precision, recall, F1-score, and ROC-AUC. A decision threshold for clinical alerting is determined by maximizing sensitivity while maintaining acceptable specificity.

Accuracy:

Acc=TP+TNTP+TN+FP+FN[36]

Precision, recall, F1:

Precision=TPTP+FPRecall=TPTP+FNF1=2⋅Precision⋅RecallPrecision+Recall[37]

ROC-AUC computed from p^ens.

Pseudo-algorithm (training & inference)

Training and inference are the two main stages of the HEALNET framework’s implementation.

The training procedure involves preprocessing multimodal IoT sensor data, encoding it using hybrid learning modules, and optimizing it using a composite loss function.

Real-time data streams are examined during inference in order to forecast cognitive states and issue notifications of anomalies.

Both procedures are summed up by the pseudo-algorithms as the following:

The system integrates multimodal sensor streams through preprocessing, feature extraction, and hybrid learning layers (CNN, LSTM, RF, SVM) with attention-based fusion, enabling personalized and privacy-preserving predictions.

Figure 3 presents the proposed HEALNET architecture. The model integrates heterogeneous sensor data streams and processes them through hybrid learning modules—CNN, LSTM, RF, and SVM—to extract temporal, spatial, and behavioural representations. A feature fusion layer combines these embeddings to produce final anomaly and decline predictions. Architecture supports adaptive learning via feedback loops and privacy-preserving mechanisms, ensuring continuous model refinement and secure-deployment. The HEALNET architecture for detecting cognitive decline in IoT-based smart home settings is depicted in Figure 3. It includes:

Input layer: multimodal IoT sensor streams (motion, door, appliance, and environmental).

• Preprocessing block: feature extraction, segmentation, norm-alization, and noise filtering.
• Feature learning modules: RF/SVM for statistical patterns, CNN for spatial features, and LSTM for temporal patterns.
• Fusion layer: a composite of learnt representations based on attention.
• Output layer: creation of anomaly alerts and classification of cognitive states.

Figure 4 illustrates the overall research workflow adopted in this study. It outlines the data-driven experimental process for behavioural analysis in smart home environments. The framework begins with multimodal IoT data acquisition, followed by preprocessing (noise removal, normalization, and segmentation), feature extraction (activity, mobility, and interaction metrics), and subsequent model training and evaluation. The framework also integrates privacy-preserving and adaptive learning modules, ensuring secure, personalized operation aligned with Saudi Vision 2030 digital health goals. The methodology ensures a seamless integration of ML and ethical safeguards for secure and responsible operation within real-world smart home setups.

Figure 4 HEALNET proposed model architecture. HEALNET, Home Environment Assisted Learning Network.

Results

The proposed HEALNET model was evaluated on the CASAS smart-home datasets using performance metrics such as accuracy, precision, recall, and F1-score. The hybrid fusion of CNN, LSTM, RF, and SVM achieved a classification accuracy of 94.2% on both datasets, representing an average improvement of 8–10% over baseline models. The confusion matrix demonstrated strong true-positive recognition of behavioural patterns consistent with MCI, indicating the effectiveness of the framework in identifying early-stage behavioural irregularities. Correlation analysis further confirmed that indicators such as abnormal sleep duration and reduced activity levels were strongly associated with elevated cognitive-risk patterns rather than clinical diagnosis. Overall, the system effectively captured temporal and spatial relationships in behavioural data, resulting in robust predictive performance and interpretability.

Experimental results

Experimental results show that HEALNET outperformed traditional classifiers, achieving 94.2% accuracy, 93.8% precision, and an F1-score of 0.92 on the test set. The confusion matrix analysis revealed a high recognition rate for MCI-consistent behavioural patterns, confirming the ability of HEALNET to identify early behavioural deviations associated with cognitive risk in older adults.

Confusion matrix

The confusion matrix presents a detailed evaluation of the classification performance of the HEALNET model across three behavioural risk categories: healthy, MCI, and cognitive decline. Based on the matrix analysis, HEALNET demonstrated high true-positive classification rates across all classes, particularly for behavioural patterns aligned with MCI risk, which are critical for early behavioural screening. The low false-positive and false-negative rates indicate that the model reliably distinguishes subtle behavioural deviations derived from smart-home activity data. This evaluation confirms that the hybrid integration of CNN, LSTM, RF, and SVM effectively captures spatial and temporal activity patterns, enabling accurate behavioural monitoring rather than clinical diagnosis in smart-home environments as stated in Figure 5.

Figure 5 Confusion matrix of HEALNET model on diverse dataset. HEALNET, Home Environment Assisted Learning Network; MCI, mild cognitive impairment.

Correlation matrix

The correlation matrix examines relationships among influential behavioural features extracted from smart-home IoT sensors. A strong positive correlation between activity frequency and activity duration indicates that more active individuals tend to exhibit more consistent daily routines. Conversely, features such as sleep irregularity and reduced activity variability show strong negative associations with cognitive-risk indicators, highlighting their relevance as behavioural signals of early decline. By identifying interdependencies among sensor-derived features, the correlation matrix supports feature selection and dimensionality reduction, enabling HEALNET to focus on the most informative behavioural attributes while minimizing redundancy during training (see Figure 6).

Figure 6 Correlation matrix of behavioural features.

Accuracy table

Table 3 provides a summary of the performance metrics of the HEALNET model. The best results were obtained with ensemble architectures consisting of CNN, LSTM, RF, and SVM.

Comparison table

The comparison in Table 4 compares HEALNET with the baseline algorithm. It highlights the superior performance of HEALNET in terms of accuracy and generalization compared to the CASAS dataset.

Comparison of existing studies

The following Table 5 compares the proposed model with existing studies on cognitive recognition in smart homes. Using multimodal sensor fusion and hybrid learning, HEALNET outpaces previous research in accuracy and feature prosperity.

Discussion

The findings demonstrate that smart-home IoT data can be effectively utilised to analyse behavioural patterns associated with cognitive health. Integrating DL with conventional machine-learning techniques improved robustness and interpretability. HEALNET successfully models sequential activity patterns, offering improved behavioural insight compared with earlier approaches limited to static sensor readings. Importantly, while the observed behavioural trends align with patterns reported in clinically characterised cohorts, the results should be interpreted as behavioural risk indicators rather than diagnostic outcomes. Challenges remain regarding data redundancy, personalisation, and privacy preservation. Future research may incorporate edge AI and federated learning to improve scalability and ethical data usage. Overall, HEALNET represents a research-stage framework for AI-assisted behavioural cognitive-risk monitoring in intelligent environments.

Accuracy graph line chart

The relative performance of many classification models—SVM, RF, LSTM, CNN, and the HEALNET model suggested here—on the CASAS smart home datasets is displayed in the accuracy line chart in Figure 7. As the models proposed which is quite different from the conventional ML models to DL hybrids, the trend line demonstrates ongoing improvement. With a maximal accuracy of 94.2%, the HEALNET curve performs noticeably better than the baseline models. This steady rising trend establishes how important hybrid models that incorporate CNN and LSTM mechanisms are for identifying notable aberrations in temporal and spatial behavioural interactions. Additionally, this figure shows modest overfitting and model stability across several cross-validation folds as in Figure 7.

Figure 7 Comparison of existing models in cognitive decline detection with proposed model (3,7,21).

Comparison of models

The key performance metrics for each algorithm employed in this study, including precision, accuracy, recall, and F1-score, are shown graphically in the model comparison graph in Figures 8,9. The advanced models like CNN and LSTM can give superior prediction accuracy since they can extract information more efficiently, but traditional models including LR and SVM perform moderately. When managing diverse IoT data streams, the proposed HEALNET model—which makes use of CNN, LSTM, and ensemble learning—performs better than other models, showing its suppleness and adaptability. Additionally, the study establishes how the This hybrid approach reduces false negatives and improves sensitivity to early behavioural deviations, enhancing its suitability for cognitive-risk screening rather than clinical diagnosis as shown in Figure 8.

Figure 8 Comparison of methods on CASAS dataset with different models. CASAS, Centre for Advanced Studies in Adaptive Systems; HEALNET, Home Environment Assisted Learning Network; LSTM, long short-term memory; SVM, support vector machine.

Figure 9 Accuracy comparison graph of proposed study with existing studies. HEALNET, Home Environment Assisted Learning Network; LSTM, long short-term memory; RF, Random Forest; SVM, support vector machine.

Comparison of existing studies

Comparative studies with previous studies further reinforce the proposed HEALNET model’s contribution to the field of cognitive health monitoring. While previous efforts—such as Dawadi et al. [2016] (7) and Akl et al. [2017] (21)—achieved accuracies ranging from 85% to 89% using traditional classifiers, new DL techniques have achieved levels of approximately 91%. However, none of these effectively incorporated multimodal behavioural data or adaptive learning processes. Comparative analysis further reinforces the contribution of HEALNET to behavioural cognitive-health research. HEALNET’s 94.2% accuracy reflects the benefit of multimodal behavioural fusion and temporal-spatial modelling, demonstrating the potential of hybrid architectures for transforming passive smart-home data into meaningful behavioural risk indicators as shown in Figures 9,10.

Figure 10 Performance comparison of models based on F1-score. HEALNET, Home Environment Assisted Learning Network; LSTM, long short-term memory; RF, Random Forest; SVM, support vector machine.

Limitations

The proposed HEALNET framework has several limitations that must be acknowledged despite its strong empirical performance. Model accuracy is largely dependent on the quality and completeness of IoT sensor data, which may be affected by inconsistent user behaviour, missing values, or sensor noise. Additionally, the datasets used in this study are limited to specific smart-home environments, namely CASAS which may restrict the generalisability of the findings across different cultural, social, and residential contexts. A key limitation of this study is the absence of clinically validated ground-truth labels, such as neuropsychological test scores or physician-confirmed diagnoses. As a result, the outputs of HEALNET should not be interpreted as clinical diagnoses but rather as behavioural indicators associated with elevated cognitive risk. Furthermore, the computational complexity of the hybrid DL architecture may pose challenges for deployment on low-power IoT or edge devices. Despite its effectiveness in identifying behavioural irregularities, HEALNET does not currently incorporate emotional, affective, or psychological signals that could provide a more comprehensive representation of cognitive health. These limitations highlight the need for larger, clinically characterised datasets, multimodal sensing strategies, and adaptive deployment mechanisms in future research.

Conclusions

This study presents HEALNET, an intelligent hybrid framework designed to analyse behavioural patterns derived from smart-home IoT data to identify behavioural indicators associated with cognitive decline. By integrating CNN, LSTM, RF, and SVM models, HEALNET effectively captures spatial and temporal dynamics in daily activity data, achieving a classification accuracy of 94.2% on publicly available datasets. The findings demonstrate that IoT-based behavioural monitoring can serve as a non-invasive and data-driven approach for cognitive-risk screening and research-stage assessment, rather than clinical diagnosis, when combined with advanced ML and DL techniques. Importantly, HEALNET is intended as a proof-of-concept framework for behavioural analysis and cognitive-risk support, not as a replacement for medical evaluation. Aligned with the objectives of Saudi Vision, this research contributes to the digital transformation of healthcare by promoting innovation in smart-home analytics and preventive monitoring. The proposed framework highlights the potential of AI and IoT technologies to support proactive, patient-centred behavioural monitoring and quality-of-life enhancement, while emphasising the need for further validation before real-world clinical adoption.

Future work may focus on extending HEALNET into a more scalable and adaptive monitoring framework by incorporating additional data modalities, such as wearable sensors, ambient audio patterns, and physiological signals, to enrich behavioural representations. Federated and privacy-preserving learning techniques may be explored to enhance data security while enabling collaborative model training across heterogeneous households. By integrating XAI, physicians will be better able to comprehend the logic’s underlying predictions, which will increase clinical suitability and confidence. Additionally, cross-cultural research using vast, diverse datasets will assurance the model’s justice and resilience across a range of demographics. Finally, the integration of edge computing technologies may reduce latency and make HEALNET more suitable for real-time intelligent monitoring in independent living environments and elderly care institutions, potentially supporting more proactive cognitive healthcare management.

Acknowledgments

The authors extend their appreciation to Prince Sattam bin Abdulaziz University for funding this research work through the project number (PSAU/2025/01/33781).

Footnote

Peer Review File: Available at https://mhealth.amegroups.com/article/view/10.21037/mhealth-2025-74/prf

Funding: The authors extend their appreciation to Prince Sattam bin Abdulaziz University for funding this research work through the project number (PSAU/2025/01/33781).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://mhealth.amegroups.com/article/view/10.21037/mhealth-2025-74/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. WHO. Dementia [Internet]. 2025 [cited 2026 Jan 7]. Available online: https://www.who.int/news-room/fact-sheets/detail/dementia
2. WHO. Anxiety disorders [Internet]. 2025 [cited 2026 Jan 7]. Available online: https://www.who.int/news-room/fact-sheets/detail/anxiety-disorders
3. Alsubai S, Alqahtani A, Sha M, et al. Smart Home-Based Complex Interwoven Activities for Cognitive Health Assessment. J Sens 2022;2022:1-10.
4. Cook DJ, Crandall AS, Thomas BL, et al. CASAS: A Smart Home in a Box. Computer (Long Beach Calif) 2013;46: [Crossref] [PubMed]
5. Ahamed F, Shahrestani S, Cheung H. Internet of Things and Machine Learning for Healthy Ageing: Identifying the Early Signs of Dementia. Sensors (Basel) 2020;20:6031. [Crossref] [PubMed]
6. Kim D, Bian H, Chang CK, et al. In-Home Monitoring Technology for Aging in Place: Scoping Review. Interact J Med Res 2022;11:e39005. [Crossref] [PubMed]
7. Dawadi PN, Cook DJ, Schmitter-Edgecombe M. Automated Cognitive Health Assessment From Smart Home-Based Behavior Data. IEEE J Biomed Health Inform 2016;20:1188-94. [Crossref] [PubMed]
8. Seelye A, Mattek N, Sharma N, et al. Weekly observations of online survey metadata obtained through home computer use allow for detection of changes in everyday cognition before transition to mild cognitive impairment. Alzheimers Dement 2018;14:187-94. [Crossref] [PubMed]
9. Bernstein JPK, Dorociak K, Mattek N, et al. Unobtrusive, in-home assessment of older adults’ everyday activities and health events: associations with cognitive performance over a brief observation period. Neuropsychol Dev Cogn B Aging Neuropsychol Cogn 2022;29:781-98. [Crossref] [PubMed]
10. Paudel R, Dunn K, Eberle W, et al. Cognitive Health Prediction on the Elderly Using Sensor Data in Smart Homes. FLAIRS 2018:317-22.
11. Seelye A, Leese MI, Dorociak K, et al. Feasibility of In-Home Sensor Monitoring to Detect Mild Cognitive Impairment in Aging Military Veterans: Prospective Observational Study. JMIR Form Res 2020;4:e16371. [Crossref] [PubMed]
12. Au-Yeung WM, Liu Y, Hanna R, et al. Feasibility of Deploying Home-Based Digital Technology, Environmental Sensors, and Web-Based Surveys for Assessing Behavioral Symptoms and Identifying Their Precipitants in Older Adults: Longitudinal, Observational Study. JMIR Form Res 2024;8:e53192. [Crossref] [PubMed]
13. Javed AR, Fahad LG, Farhan AA, et al. Automated cognitive health assessment in smart homes using machine learning. Sustain Cities Soc 2021;65:102572.
14. Shahid ZK, Saguna S, Åhlund C. Forecasting and Detecting Anomalies in ADLs in Single-Resident Elderly Smart Homes. Proceedings of the International Conference on Research in Adaptive and Convergent Systems. New York, NY, USA: ACM; 2023:1-8.
15. Gupta BB, Gaurav A, Attar RW, et al. LSTM Based Neural Network Model for Anomaly Event Detection in Care-Independent Smart Homes. Computer Modeling in Engineering & Sciences 2024;140:2689-706.
16. Bokhari MU, Yadav G, Zeyauddin AS. Enhancing mental health prognosis: an investigation of advanced hybrid classifiers with cutting-edge feature engineering and fusion strategies. International Journal of Information Technology. 2024; [Crossref]
17. Chimamiwa G, Giaretta A, Alirezaie M, et al. Are Smart Homes Adequate for Older Adults with Dementia? Sensors (Basel) 2022;22:4254. [Crossref] [PubMed]
18. Chung J, Pretzer-Aboff I, Parsons P, et al. Using a Device-Free Wi-Fi Sensing System to Assess Daily Activities and Mobility in Low-Income Older Adults: Protocol for a Feasibility Study. JMIR Res Protoc 2024;13:e53447. [Crossref] [PubMed]
19. Alsubai S, Alqahtani A, Sha M, et al. Automated Cognitive Health Assessment Based on Daily Life Functional Activities. Comput Intell Neurosci 2023;2023:5684914. [Crossref] [PubMed]
20. Yadav G, Bokhari MU. Hybrid Classifier for Optimizing Mental Health Prediction: Feature Engineering and Fusion Technique. Int J Ment Health Addict 2024; [Crossref]
21. Akl A, Snoek J, Mihailidis A. Unobtrusive detection of mild cognitive impairment in older adults through home monitoring. IEEE J Biomed Health Inform 2017;21:339-48. [Crossref] [PubMed]
22. Grammatikopoulou M, Lazarou I, Alepopoulos V, et al. Assessing the cognitive decline of people in the spectrum of AD by monitoring their activities of daily living in an IoT-enabled smart home environment: a cross-sectional pilot study. Front Aging Neurosci 2024;16:1375131. [Crossref] [PubMed]
23. Band SS, Yarahmadi A, Hsu CC, et al. Application of explainable artificial intelligence in medical health: A systematic review of interpretability methods. Inform Med Unlocked 2023;40:101286.
24. Austin J, Klein K, Mattek N, et al. Variability in medication taking is associated with cognitive performance in nondemented older adults. Alzheimers Dement (Amst) 2017;6:210-3. [Crossref] [PubMed]
25. Pierson E, Althoff T, Leskovec J. Modeling Individual Cyclic Variation in Human Behavior. Proc Int World Wide Web Conf 2018;2018:107-16.
26. Tiersen F, Batey P, Harrison MJC, et al. Smart Home Sensing and Monitoring in Households With Dementia: User-Centered Design Approach. JMIR Aging 2021;4:e27047. [Crossref] [PubMed]
27. Gettel CJ, Chen K, Goldberg EM. Dementia Care, Fall Detection, and Ambient-Assisted Living Technologies Help Older Adults Age in Place: A Scoping Review. J Appl Gerontol 2021;40:1893-902. [Crossref] [PubMed]
28. Xu Z, Vekaria V, Wang F, et al. Using Machine Learning to Predict Antidepressant Treatment Outcome From Electronic Health Records. Psychiatr Res Clin Pract 2023;5:118-25. [Crossref] [PubMed]
29. Hu Y, Chen J, Chen J, et al. An Ensemble Classification Model for Depression Based on Wearable Device Sleep Data. IEEE J Biomed Health Inform 2024;28:2602-12. [Crossref] [PubMed]
30. Dara OA, Lopez-Guede JM, Raheem HI, et al. Alzheimer’s Disease Diagnosis Using Machine Learning: A Survey. Applied Sciences 2023;13:8298doi: 10.3390/app13148298.
31. Vizitiu C, Stara V, Antognoli L, et al. An IoT-based cognitive impairment detection device: A newly proposed method in older adults care-choice reaction time-device development and data-driven validation. Digit Health 2024;10:20552076241293597. [Crossref] [PubMed]
32. Yadav G, Ubaidullah Bokhari M, Alzahrani SI, et al. Emotion-Aware Ensemble Learning (EAEL): Revolutionizing Mental Health Diagnosis of Corporate Professionals via Intelligent Integration of Multi-Modal Data Sources and Ensemble Techniques. IEEE Access 2025;13:11494-516.
33. Sharma SK, Alutaibi AI, Khan AR, et al. Early detection of mental health disorders using machine learning models using behavioral and voice data analysis. Sci Rep 2025;15:16518. [Crossref] [PubMed]
34. Center for Advanced Studies in Adaptive Systems (CASAS). CASAS Smart Home Data Sets [Internet]. [cited 2026 Jan 7]. Available online: https://casas.wsu.edu/datasets/