This review identifies studies that focus on ADL recognition using wearable sensors to recognize basic ADLs as defined by Katz et al [5]. A subgoal within this review is to identify studies that include older adults in the research, for example, as participants in data collection for training ADL recognition systems or in user studies centered on using wearable sensors for ADL recognition. This work is a scoping review that follows the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines [42] and the protocol [43] is registered with the Open Science Foundation. Following standard procedures for PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses), publications are systematically and hierarchically screened and assessed for eligibility. Title and abstracts are screened during the first step of the screening phase because some of our inclusion criteria (Table 1) may not appear in the title.

The focus of this survey is relevant to both the medical and computer science literature. We queried both PubMed and the Association of Computing Machinery Digital Library (ACM DL) to survey a substantial corpus of studies in both domains. In addition, we supplemented the queries with additional studies identified from Google Scholar. The queries were conducted in January 2025.

The survey summarizes the state of basic ADL recognition by covering the last 5 calendar years (2019-2024). This timeframe focuses on recent trends as technology in this area can rapidly change, for example, the proliferation of consumer smart devices.

The search terms used in the queries are provided in Table 2. These terms align with previous reviews focused on ADL recognition [37,41]. Each row in Table 2 is combined with AND Boolean logic. “Wearable” is the only term used to specify the sensors used to minimize the capture of activity recognition systems using ambient or environmental sensors. The term “basic” is not clarified in the ADL-related terms because many works do not specify basic versus instrumental.

Katz et al [5] defines BADLs as Bathing, Dressing, Toileting, Transferring, Continence, and Feeding. The ability to perform these activities serves as a score of patient independence. The Katz Index of Independence in Activities of Daily Living gives clear definitions of what skills the patient needs to have and what types of assistance they can receive; for example, food preparation is not part of the Feeding activity. In practice, Katz BADLs are treated as categories and will include related and proxy activities. For example, Bathing includes other personal hygiene activities such as brushing teeth, and Toileting includes activities such as flushing which serve as a proxy indicator of the main activity. Continence is generally combined with Toileting as the former is not an activity within the scope of activity recognition research. Table 3 provides the adjusted definitions of each BADL to align with the practices in the ADL recognition literature.

Activity recognition has a distinct literature related to ambulation activities such as walking, running, and ascending or descending stairs. These studies are diverse with respect to focusing on clinical applications, for example, remote monitoring of rehabilitation after injury, or nonclinical applications, such as exercise tracking. Many of these studies model ambulation as different states of being (eg, “the person is currently sitting” or “the person is currently walking”) and do not capture transitions such as standing up from sitting. Katz’s definition of transferring focuses on transitions rather than the patient’s general mobility; we do not include general ambulation activities in Table 3. Recognition of ambulatory activities and related gait attributes merit their own study.

The eligibility and exclusion criteria are given in Table 1. ADL recognition needs to be a direct outcome or goal of the study to be included. Furthermore, at least one BADL needs to be included; studies focused on instrumental ADLs such as cooking were excluded. Two related areas to BADL recognition are physical activity recognition and fall detection. Physical activity recognition papers tend to focus on maintaining one’s physical health and tracking exercise, which are disparate goals from basic ADL monitoring. Some studies in fall detection have the goal of distinguishing falls versus normal, daily activities, but the identification of daily activities is a strategy to decrease false positives in lieu of a direct goal of the work. As such, studies focused on general physical activity or fall detection are excluded.

The reviewing process was conducted predominantly by one author. The author met with at least 2 other authors to discuss ambiguous papers at the end of each step during the screening process. The eligibility and exclusion criteria were not subjective; one reviewer was sufficient for most papers.

The search resulted in 695 studies. After removing duplicates, the titles and abstracts of 690 publications were screened. A total of 164 studies were eligible for full-text assessment; 58 studies satisfied the inclusion criteria. Overall, 8 of these studies included older adults in the research. The paper selection flowchart is given in Figure 1.

We summarize the state-of-the-art for the recognition of BADLs in the following subsections (Table 4). We then highlight the papers that included older adults (Table 5).

Most studies use commodity devices such as smartphones and smartwatches for motion sensors and microphones regardless of which ADLs were targeted. Using these devices maximizes the solution’s potential for adoption as these devices are ubiquitous. Research using smartphones assumes that the phone will either be in the user's pocket or nearby them. Some studies strap the smartphone to a person’s arm to simulate a watch-like form factor [44] or to keep the device near the user in an unobtrusive manner [45]. Some studies still leverage custom sensor arrays, for example, the Opportunity dataset where participants had accelerometers placed all over the body [46] and the study by Bedri et al [47], which placed sensors on a pair of glasses, on the ear, and on the back of the neck.

Bathing and showering behaviors are rare inclusions for ADL recognition systems. Previous surveys did not capture any papers with these activities [37,41]. Liang et al [45] used an audio-based approach to detect sounds associated with bathtubs and showers, for example, filling with water or washing. The lack of focus on these activities has several likely reasons. First, these activities are difficult to simulate in laboratory settings due to their facility requirements. Second, wearable sensors need to be sufficiently waterproof or protected during these activities. Finally, people are more likely to be uncomfortable performing these activities while being recorded. The protocol in the study by Liang et al [45] avoided these issues. They conducted a free-living study where the researcher followed the participant around their own home, and the researcher kept a distance to minimize any influence on how activities were performed.

Many hygiene activities primarily involve the hands and can be captured with wrist-worn devices such as smartwatches. The most commonly included activities are brushing teeth, combing hair, and washing hands. Each of these activities tends to be performed for a sustained amount of time and has rhythmic attributes in their motions, making them distinct from other daily activities. Furthermore, these activities have low expected variation in their performance, making it easier for an ADL recognition model to generalize. These activities are usually included in a large set of general ADLs (ie, the work is not specifically focused on bathroom or hygiene activities). Examples of this practice include studies by Bhattacharya et al [10] and Cherian et al [22] that recognized 23 and 8 different activities, respectively, and covered a range of everyday activities. Exceptions to this pattern include Akther et al [17] who focused on assessing how thoroughly the user brushed their teeth and Mondol et al [51] and Santos-Gago et al [53] whose work focused on identifying hand-washing behaviors that are compliant with the World Health Organization (WHO) guidelines.

Dressing is an uncommon activity to be included within the constraints of this review. Motion sensors such as accelerometers are the most common type of sensor used in wearable systems. However, the motions associated with dressing activities tend to be subtle, making them difficult to distinguish from other everyday activities. Furthermore, the diversity in the styles of clothing people wear makes designing a general recognition system complicated. Sun et al [15] included “putting on clothes” and “taking off clothes” among the total of 221 activities included in their work focused on developing a multimodal activity recognition system. This work found that both motion and Wi-Fi signals were useful in determining if the user was interacting with clothing. Dressing in Narkhede et al [64] used both motion and location data and commented that location context was necessary due to the high variability in the motion data.

Toileting behaviors are generally detected via a proxy indicator. Using a toilet or urinal does not involve significant bodily motion, therefore common locations for wearable sensors will not detect these activities. However, this activity is normally followed by flushing the toilet or urinal, as this sound is used to signal the end of the activity. Liang et al [45] use only sound, Masum et al [65] uses only motion, and Mollyn et al [59] uses both. Liang et al [45] and Mollyn et al [59] specify that their systems recognize the flushing action while Masum et al [65] focuses on the action of sitting on the toilet.

Ambulation activities are widely studied in activity recognition research.

Many studies developed methods for distinguishing between different modes of ambulation and posture such as walking, jogging, ascending or descending stairs, sitting, standing, and lying down [88-93]. However, most of these studies focus on just detecting the current state of the user and do not directly capture the transitions between modes (eg, sit-to-stand, stand-to-sit, lying-to-stand, or stand-to-lying). The datasets UniMiB-SHAR [32], MobiAct [30], and Transitional Activities [94] do include these annotations, and many recent deep learning–centric studies have leveraged these datasets to detect these activities [11,24-29]. The first 2 datasets involve a smartphone placed in the front pocket of the user’s pants while the last dataset places sensor nodes across the body on the waist, right wrist, left wrist, right arm, left thigh, and right ankle.

Eating and drinking have the most diversity in the sensors used to recognize these activities. While they can be recognized with wrist-worn sensors [10,22,77], these activities also afford opportunities to place sensors on the head or neck [47,70,85]. Bedri et al [70] used a glasses form factor to detect episodes of eating and drinking. This approach allows the system to avoid ambiguity with other hand-centric activities and focus on detecting actions such as chewing and swallowing. Some studies will use the high-level labels in the Opportunity dataset [46] to recognize examples of eating and drinking. Opportunity contains kitchen-centric activities where the participant goes through a gauntlet of activities including opening and closing cabinets, opening and closing a refrigerator door, making a sandwich, and making coffee. This dataset has low-level annotations focused on the specific motions, for example, “open drawer” and high-level annotations focused on the activity being performed, such as “sandwich time.” Some studies only focus on recognizing the 17 gestures in an activity-agnostic fashion, but others use high-level annotations to recognize actions such as taking a bite of a sandwich or taking a sip of coffee [24,58,74,83].

We highlight studies that include older adults in the main research (Table 5). Only 5 of these studies have the design and evaluation of activity recognition systems as their main contribution. The other 3 studies are highly connected to ADL recognition using wearable sensors, meriting their inclusion.

Alam et al [36] represents the concern that systems trained on younger populations will not generalize to older populations when deployed. Their work focuses on mitigating biases in their developed ADL recognition system to be robust to differences in physical ability. The main evaluation is a gesture recognition system to distinguish between 8 hand gestures and a walking recognition system that is robust to the user having a mobility aid such as a walker.

Cao et al [56] focuses on recognizing the activity of washing hands in older adults with dementia in a user-independent fashion. Their system identifies specific steps in the handwashing process to identify if the patient needs assistance in properly washing their hands. Their evaluation included 8 older adults with cognitive impairment as tested by the Montreal Cognitive Assessment (MoCA) [98].

Cook et al [57] leverages an ADL recognition system to label the participants’ behaviors in free-living conditions. The goal of this work was to distinguish between brain health intervention and nonintervention participants. The behavior predictions were used as input to the brain health intervention versus nonintervention classifier.

Kim et al [95] created a speech-based smartwatch application to gather in-situ annotations of daily activities. The design of the system focused on the needs and comfort of older adults to minimize the burden associated with data annotation. They conducted a user study over the course of 7 days with 13 older adults to evaluate the experience of using such a system. An envisioned goal for this system is facilitating personalized activity recognition.

Caldeira et al [96] and Cherian et al [97] include perspectives from older adults regarding the usage of monitoring technology for ADL performance. Calderia et al [96] interviewed participants after living in a smart home and wearing a smartwatch for 2.5 years. They found that participants wanted to be included and to engage with their data, especially with respect to determining if they were living an active lifestyle. Cherian et al [97] interviewed participants who lived in the assisted living section of a continuing care retirement community before and after wearing smartwatches for 1 week to simulate an ADL monitoring system. Participants acknowledged the potential utility of such a system and voiced a desire to maintain their independence.

Wellnitz et al [85] and Alevizaki et al [61] discuss that their systems have utility for ADL tracking for older adults and include at least one older adult in their data collection and system evaluation. However, the majority of their participants are younger adults, and including older adults was not an explicit goal.

BADLs are not equally covered in the ADL recognition literature. Few studies attempt to recognize bathing, dressing, and toileting. These activities have attributes that make them more difficult to detect as stated in the Results section. Bathing and toileting relied on sound cues, and dressing relied on location context. By contrast, the other BADLs were recognizable using motion sensors, which are ubiquitous in commercially available wearable devices. The special facility requirements and concerns for subject privacy of bathing, toileting, and dressing may have caused them to be deprioritized in the ADL recognition literature because systems that recognize many activities (ie, multiclass recognition) are a current focus. However, recognizing bathing, dressing, and toileting has value in health monitoring applications as they give medical practitioners insight into their patients’ health, for example, whether the patient is maintaining regular bathroom habits. Robustly detecting these activities is an open problem for ADL recognition researchers to address as the field progresses.

All the papers collected in this scoping review discuss older adults at some point in their work by the nature of our query. However, our results show that only 14% (8/58) of studies in this scoping review include older adults in the design and evaluation of their systems. The main reason for this disconnect is that many publicly available datasets for ADL recognition include only younger adults [30-33,46]. Many techniques in artificial intelligence and machine learning require annotated datasets, and researchers are strongly encouraged to use existing datasets to benchmark their contributions and increase the reproducibility of their work. Including older adults requires conducting a custom user study and annotating the data which is time-consuming and costly. Evaluating the system with the target audience is not a necessary condition for the completion of studies that focus on developing new systems or techniques, for example, exploring self-supervised learning. This type of research consequently remains relatively unexplored, creating a knowledge gap in how well systems trained on younger populations generalize to older populations.

Activity recognition research is united by common goals. Techniques and approaches vary greatly across works, making surveys valuable summaries of the myriad of explored solutions. The most related scoping reviews to this work are the Camp et al survey of tools for ADL tracking for community-dwelling older adults [37] and Zhang et al survey of recognition of bathroom activities with wearable devices [41]. Camp et al did not focus on the sensing medium (ie, wearable versus ambient) but instead centered on commercially available devices. The study scope of Zhang et al was limited to only bathroom activities which are understudied as supported by this work. The current contribution is distinct by focusing on summarizing the state-of-the-art research approaches for all BADLs.

One of the goals of this work was to summarize the current practices in the design of ADL recognition systems for BADLs. A limitation of the search strategy used in this review is the inability to establish a performance benchmark for future research. We summarize the common techniques and considerations the studies contribute, but the performance evaluations of their solutions were not captured due to the heterogeneity of the data.

Another limitation is that our approach does not provide concrete insights into how to incorporate wearable ADL tracking systems into aging-in-place solutions. The studies in this survey predominantly focus on proving system feasibility, not usability. Understanding what older adults, their caregivers, and their loved ones desire in an ADL tracking system for supporting aging in place is an avenue for future research.

Our work focused on ADL tracking systems for BADLs to support aging in place. Future surveys should consider systems for recognizing instrumental ADLs or other solutions for supporting aging in place. Potential avenues include intelligent systems for home automation and social robotics for assistance in performing ADLs.

Most systems in activity recognition are evaluated in terms of correctly identifying the activity performed. However, this problem definition ignores the additional information about how the activity was performed (ie, whether it was performed correctly, adequately, or even abnormally). Systems that can detect and assess activity performance have more utility for caregivers and family members, providing more insight into the individual's health status. Specifically, knowing the quality of the activity performance can be used to detect declines in health and signal the need for an intervention or increase in needed care. Designing ADL recognition systems that can detect degradation in ADL performance is a potential avenue for future work.

Over the coming decades, the population of older adults is expected to increase significantly, a trend that will put tremendous strain on health care systems around the world. Making it possible for people to stay in their homes longer safely (ie, aging in place) has great potential clinical implications. One solution to support aging in place is using human activity recognition systems to automatically track ADL performance, providing a safety net that can detect significant changes in ADL performance.

Wearable ADL recognition has promise for enabling these aging-in-place systems, but the current literature has several gaps to be addressed before this option becomes feasible. Several basic ADLs (eg, bathing, dressing, and toileting) have little coverage and remain open problems for ADL recognition. Additionally, many works are predominantly concerned with proving system feasibility and do not assess usability or real-world deployment. For these systems to move from academic experiments to actual systems with clinical utility, ADL recognition systems must consider the design requirements of being part of remote health monitoring systems.

In this survey, we reviewed human activity recognition systems designed to recognize basic ADLs using wearable sensors. Despite targeting older adults as users, many studies do not directly include this population in their research. To address this gap, ADL recognition researchers are encouraged to evaluate their systems with older adults as participants to assess how their systems would work in a real-world deployment.