An Intelligent Non-Invasive Real Time Human Activity Recognition System for Next-Generation Healthcare

1. Introduction

• Remote monitoring is framed as a way for vulnerable people to live more independently while enabling caregivers to receive real-time fall messages.The introduction connects this motivation to the growing elderly population and limited caregiver capacity.Source: §1 · pp. 1–2
• The study proposes a USRP-based dataset and machine-learning classification of sitting versus standing, alongside comparison with a wearable-device benchmark.The introduction identifies these as the paper’s two main contributions.Source: §1 · p. 2
• Wearables can support monitoring but may be uncomfortable, forgotten, or raise concerns about continuous tracking.Radio-wave sensing is presented as a non-invasive alternative using changes in channel state information (CSI).Source: §1 · p. 2

2. Related Work

• 91.5% was the highest result reported for wearable-accelerometer activity datasets.The reviewed work compared Random Forest, KNN, and SVM.Source: §2 · pp. 2–3
• 95.5% accuracy was reported using KNN on FMCW radar data from three participants performing falls and related activities.The passage also notes multichannel extraction work.Source: §2 · p. 3
• 85.17% average accuracy was reported for a 14-activity wearable-sensor benchmark, with 92.16% for static and 80.0% for dynamic activities.A separate smartwatch study of ping-pong motions reported 97.80% accuracy with Random Forest.Source: §2 · p. 3
• 96% accuracy was reported for CSI-based classification of five arm movements using LSTM.Other reviewed work used CSI to identify people, with higher scores in binary experiments involving two people.Source: §2 · p. 3

3. COLLECTION OF DATA

• Two USRPs—an X300 transmitter and X310 receiver—connected to PCs used VERT2450 omnidirectional antennas and operated four metres apart in line of sight in an office environment.The simulation used MATLAB/Simulink, and the office space measured 7 by 8 metres.Source: §3 · p. 3
• The dataset contains 30 samples each of sitting and standing, recorded as CSI while participants moved between the antennas.The 64-subcarrier traces show frequency on the Y-axis and time on the X-axis, with each color representing a subcarrier; Figures 1 and 2 show good samples for the respective motions.Source: §3 · p. 4, Figure 1 · p. 5, Figure 2 · p. 6
• USRPs transmitted for 10 seconds per sample, and packet counts varied slightly between recordings without materially affecting the pattern-detection aim.Repeated samples were collected because motion and signal propagation varied, and some recordings could be affected by ambient movement or atmosphere.Source: §3 · pp. 4, 6
• The experiment-flow-chart visual is identified as Figure 3, while Tables 1 and 2 are identified as software- and hardware-configuration parameter selections.The supplied text names these visuals but does not specify their detailed contents.Source: Table 1 · p. 4, Table 2 · p. 5, Figure 3 · p. 6

4. Machine Learning Process

• Samples are converted to CSV, processed with Pandas, and combined; missing values from shorter samples are replaced with zeros before classification.The authors state that the shorter samples’ trailing zeros retain the radio-signal pattern.Source: §4 · pp. 6–7
• Random Forest combines decision-tree predictions by majority vote, with each tree using features learned during training.The section gives the weighted impurity-based expression used by SciKit for Random Forest node importance.Source: §4 · p. 7
• KNN assigns test data to the nearest matching training samples using the default Euclidean-distance method.The passage describes K as the number of samples and identifies x as data and y as the label.Source: §4 · p. 7
• SVM trains hyperplanes from training-data features to separate inputs into categories.The section also presents an inequality for the SVM and identifies its vector, unknown-vector, and constraint terms.Source: §4 · p. 7
• The neural network adjusts hidden-layer weights when training outputs are incorrect, then uses test inputs to produce predictions.Its architecture comprises interconnected input, hidden, and output layers.Source: §4 · p. 8
• Each algorithm is tested with 10-fold cross-validation and a 70/30 train-test split, using accuracy, precision, recall, and F1-score.Precision and recall are averaged across sitting and standing; F1-score averages precision and recall.Source: §4 · p. 8

5. Results and Discussion

• The results section evaluates accuracy, precision, recall, and F1 score, and uses confusion matrices to show sample classifications.The cross-validation and train-test experiments use different evaluation procedures.Source: §5 · pp. 8–9

5.1. Cross Validation

• Under 10-fold cross-validation, Random Forest ranked first and Neural Network second; ensemble accuracy was 92.18%, KNN scored 88.17%, and SVM 84.68%.The 10 test-fold results were averaged.Source: §5.1 · pp. 9–10
• Random Forest correctly classified 1,821 sitting and 1,730 standing samples in cross-validation, with 99 sitting and 190 standing samples misclassified.Figure 4 presents the confusion matrix for this classifier.Source: Figure 4 · p. 9, §5.1 · pp. 9–10

5.2. Train Test Split

• With a 70/30 train-test split, Random Forest ranked first; the ensemble reached 93.83% accuracy, Neural Network 93.40%, KNN 90.71%, and SVM 81.77%.The test set contained 1,152 samples from the 3,840-subcarrier dataset.Source: Table 4 · p. 12, §5.2 · pp. 12–14
• In the train-test experiment, Random Forest correctly classified 606 sitting and 508 standing samples; 34 sitting and 4 standing samples were misclassified.Figure 9 shows the confusion matrix for the 1,152 test samples.Source: Figure 9 · p. 12, §5.2 · pp. 12–13

5.3. Comparison of Cross Validation and Train Test Split

• Most algorithms had higher accuracy with the train-test split except SVM, while cross-validation was described as more representative because all data take turns in training and testing.The authors attribute stronger standing-sample classification in the split experiment to more standing samples in the training set.Source: §5.3 · p. 15, Figure 14 · p. 15

5.4. Real Time classification

• The Flask interface connects Python to Matlab-recorded CSI, applies a saved Random Forest model, and correctly classified all 12 unseen sitting and standing samples.The test comprised six samples of each activity.Source: §5.4 · pp. 15–16

5.5. Benchmark dataset

• USRP results were similar to the wearable benchmark; Random Forest performed best on both datasets, while the ensemble slightly favored the benchmark under cross-validation.The comparison also reports small differences for several algorithms, with the direction varying by algorithm and evaluation method.Source: Table 5 · p. 17, Table 6 · p. 17, §5.5 · p. 18

• 92.47 % accuracy was achieved by Random Forest on the proposed dataset.Source: §6 · p. 18
• The dataset records USRP channel-state information for binary classification of sitting down versus standing up.Source: §6 · p. 18
• The web application successfully classified samples absent from the learning phase using CSI from the USRP.Source: §6 · pp. 18–19
• The wearable-device benchmark also showed good accuracy when evaluated with the same machine-learning techniques.Source: §6 · p. 19