Software Design

Machine Learning and Interactive Game Interface

We used a binary logistic classification algorithm to classify triceps sEMG signals. The primary objective of the model is to distinguish between relaxed and active muscle states, enabling the system to determine whether the user is intentionally activating their triceps to trigger in-game actions and interact with on-screen elements. Specifically, the model classifies real-time sEMG signals into two labels: Relaxed (0) and Active Contraction (1).

We chose a Logistic Regression model for the following reasons:

We extract four commonly used EMG features:

These four features effectively capture signal intensity, amplitude, and waveform characteristics, providing reliable separation between relaxed and active muscle states. Additionally, we implemented a separate module for the stretchable respiration sensor to support deep-breathing training. This module continuously estimates the user's breathing state and controls the behaviour of a training ball within the game environment.

For sEMG signal acquisition, preprocessing is essential because sEMG is a weak biosignal and may become saturated without proper signal conditioning. To reduce noise and prevent possible saturation, we applied a low-pass filter in addition to the hardware-side amplification system. Data were collected using the Processing IDE via a simple serial communication protocol. The collected data were saved as CSV files and subsequently processed in IntelliJ IDEA. For model training, datasets were collected from four participants. To obtain accurately labeled datasets, the training data were recorded such that the first 0–30 s contained stable resting-state signals, while the 30–60 s interval contained active muscle contraction signals. We labeled the first 1500 samples as class 0 (Relaxed) and the remaining samples as class 1 (Active Contraction). After labeling, the data were segmented into windows, the four EMG features were extracted, and the corresponding labels were assigned to each window. These extracted features were then used to train the logistic regression classifier, which was exported as a .pkl file containing the feature scaler (mean and variance), model coefficients, all model parameters, and the bias term. This allows the model to be loaded directly or reimplemented in Java for real-time deployment.