AI Fills Gaps in Human Activity Data

Missing data from smartphones and wearables can undermine apps that track daily activities like walking or running, affecting everything from fitness monitoring to elderly care. A new study addresses this by using artificial intelligence to intelligently fill in these gaps, ensuring more reliable recognition of human movements without needing extra sensors or invasive data collection.

The researchers developed a technique that identifies and reconstructs missing samples in activity datasets. By applying the k-Nearest Neighbors (KNN) algorithm, the system estimates absent data points based on similar, existing patterns. This approach avoids fabricating random values, instead using real data relationships to maintain accuracy in identifying activities such as walking, running, standing, moving upstairs, and moving downstairs.

The methodology starts with data acquisition from publicly available datasets, which include recordings from accelerometers, gyroscopes, and magnetometers in mobile devices. These sensors capture motion data, but issues like low battery or memory constraints can lead to incomplete records. The process involves four stages: first, identifying missing samples by checking data frequency; second, inserting null values to mark gaps; third, segmenting data into chunks for analysis; and finally, using KNN to impute the missing values by finding the closest matches in the dataset.

Results from testing on a dataset with activities recorded over 5-second intervals showed significant improvements. For example, in a moving downstairs activity, the accelerometer data originally had 125 missing samples out of a required 500. After imputation, the reconstructed data closely matched the high-amplitude patterns typical of real movements, as illustrated in the paper's figures. This method proved effective in maintaining data integrity, with the imputed samples allowing for better feature extraction in machine learning models used for activity recognition.

This advancement matters because it enhances the reliability of applications in health monitoring, sports medicine, and assisted living. For instance, fitness apps can provide more accurate feedback on exercises, while systems for older adults can better detect falls or daily routines without requiring constant sensor uptime. By improving data quality, the technique supports broader adoption of AI in everyday technology, helping devices work smarter with existing hardware.

Limitations noted in the study include the method's dependency on having enough similar data points for accurate imputation. If too many samples are missing, the approach may struggle, and in cases where data is severely incomplete, files might need to be discarded. Future work aims to refine the imputation to handle varying data amplitudes and integrate it into real-time systems for even better pattern detection.