A systematic review of smartphone-based human activity recognition for health research

Background: Smartphones are now nearly ubiquitous; their numerous built-in sensors enable continuous measurement of activities of daily living, making them especially well-suited for health research. Researchers have proposed various human activity recognition (HAR) systems aimed at translating measurements from smartphones into various types of physical activity. In this review, we summarize the existing approaches to smartphone-based HAR. Methods: We systematically searched Scopus, PubMed, and Web of Science for peer-reviewed articles published up to December 2020 on the use of smartphones for HAR. We extracted information on smartphone body location, sensors, and physical activity types studied and the data transformation techniques and classification schemes used for activity recognition. Results: We identified 108 articles and described the various approaches used for data acquisition, data preprocessing, feature extraction, and activity classification, identifying the most common practices and their alternatives. Conclusions: Smartphones are well-suited for HAR research in the health sciences. For population-level impact, future studies should focus on improving quality of collected data, address missing data, incorporate more diverse participants and activities, relax requirements about phone placement, provide more complete documentation on study participants, and share the source code of the implemented methods and algorithms.

💡 Research Summary

This systematic review collates and critically evaluates the state of smartphone‑based human activity recognition (HAR) as it pertains to health research. By searching Scopus, PubMed, and Web of Science up to December 2020, the authors identified 108 peer‑reviewed studies that employed consumer‑grade smartphones to capture, process, and classify everyday physical activities. The review first maps the hardware landscape: the tri‑axial accelerometer and gyroscope are the dominant sensors, often complemented by magnetometers or barometers in a minority of works. Device placement varies widely—pocket, waist, chest, and wrist are the most common sites—yet the authors note that placement strongly influences signal characteristics, necessitating careful preprocessing such as gravity removal, coordinate alignment, and low‑pass filtering (typically 0.25–20 Hz).

Data segmentation follows a largely standardized protocol: sliding windows of 2–5 seconds with 50 % overlap are used to balance temporal resolution and computational load. Feature extraction is split between time‑domain descriptors (mean, standard deviation, RMS, zero‑crossing rate, etc.) and frequency‑domain descriptors (FFT‑based power spectra, Mel‑frequency cepstral coefficients, wavelet coefficients). While traditional machine‑learning pipelines rely on these handcrafted features, the review highlights a growing shift toward deep learning models that learn representations directly from raw or minimally processed sensor streams.

Classification algorithms are catalogued in detail. Support Vector Machines, Random Forests, and k‑Nearest Neighbors remain the workhorses of the field, achieving accuracies between 80 % and 95 % for simple activities such as walking, sitting, or stair climbing. More complex or mixed activities are increasingly tackled with Convolutional Neural Networks (CNN), Long Short‑Term Memory networks (LSTM), and hybrid CNN‑LSTM architectures, which consistently outperform conventional methods in recent studies.

Despite these technical advances, the authors identify several systemic shortcomings that limit real‑world impact. Sample sizes are modest, and demographic diversity (age, sex, body composition, cultural context) is often insufficient, raising concerns about model generalizability. Rigid requirements on phone placement reduce ecological validity, as users in everyday life rarely adhere to prescribed positions. Handling of missing data, sensor dropout, and noise is inconsistently reported, undermining reproducibility. Moreover, few papers make their code or datasets publicly available, impeding benchmarking and collaborative progress.

To address these gaps, the review proposes a forward‑looking agenda: (1) develop large, openly shared datasets that capture a broad spectrum of participants and activities; (2) design algorithms robust to arbitrary device placement, possibly through sensor‑fusion or adaptive calibration techniques; (3) standardize missing‑data imputation and sensor‑fault detection methods; (4) mandate transparent reporting of participant characteristics, preprocessing pipelines, and hyper‑parameter settings; and (5) encourage open‑source release of code and trained models.

In conclusion, smartphones—owing to their ubiquity, low cost, and rich sensor suites—are uniquely positioned to enable scalable, continuous monitoring of human activity for health research. By embracing the recommended best practices, future work can translate methodological innovations into actionable insights for disease prevention, lifestyle interventions, and remote patient monitoring, thereby realizing the full public‑health potential of smartphone‑based HAR.