BrainFusion: a Low‐Code, Reproducible, and Deployable Software Framework for Multimodal Brain‒Computer Interface and Brain‒Body Interaction Research

Figures

• 

  Figure 1. Overview of BrainFusion.

• 

  Figure 2. Multimodal physiological signal processing workflow in BrainFusion. Raw data from EEG, fNIRS, EMG, and ECG are organized in a standardized data container. Signals undergo modality‐specific preprocessing, followed by feature extraction across multiple analytical domains through time‐domain (TDA), frequency‐domain (FDA), and time‐frequency (TFDA) analysis, along with network and nonlinear methods. Dynamic coupling metrics such as NVC, CMC, and HBI are calculated and visualized through the multimodal dynamic coupling analysis engine (MDCAE).

• 

  Figure 3. Illustration of the three major functional modules in BrainFusion. A) The one‐click workflow designer includes the UI layer for drag‐and‐drop workflow configuration, the connection layer for ensuring DAG structure compliance and node connection validation, and the execution layer comprising modular functional nodes. B) The machine learning extension framework supports traditional AutoML‐based model selection, deep learning benchmarking, and integration of user‐defined models via the external model interface. C) The application generator converts workflows into standalone executable applications with a GUI, enabling rapid deployment and demonstration.

• 

  Figure 4. Analysis of the TU‐Berlin BCI dataset via BrainFusion. A) Comparison of EEG and fNIRS data before and after preprocessing. B) Topographic maps of EEG frequency bands (δ, θ, α, β, γ) during left‐handed and right‐handed motor imagery, and rest. C) The fNIRS channel layout corresponds to the motor cortex (C3/C4 regions), with average HbO and HbR responses extracted for motor‐related hemodynamic analysis. D) Neurovascular coupling modeling: EEG‐derived PSD (C3) convolved with a canonical HRF to predict fNIRS HbO signals. E) Left: Channel layout showing selected EEG electrodes and fNIRS channels. Right: Comparison of static NVC strengths between left and right‐hand motor tasks across EEG channels. Box plots show the distribution of mean absolute NVC values (range: 0–1) for each channel under left‐hand (light blue) and right‐hand (dark blue) conditions. Data represent 870 paired samples. No statistically significant differences were observed between the two conditions for any channel (two‐tailed paired t‐test, p < 0.05). F) Motor imagery classification performance. Violin plots showing the per‐subject mean 10‐fold cross‐validation accuracies (distribution), and bar plots showing the…

• 

  Figure 5. The results of sleep staging classification and generated application. A) Performance comparison of four deep learning models on the HMC dataset, presenting mean ± standard deviation (SD) for overall accuracy, macro‐averaged F1 score (mF1), and individual F1 scores across sleep stages (Wake, N1, N2, N3, REM). All metrics are averaged over 5‐fold cross‐validation. Error bars indicate one SD from the mean. B) User interface of the prototype inference application generated based on CareSleepNet, supporting real‐time sleep stage prediction, probability score visualization, and historical inference tracking.