Dapalm | IMS技术前沿

NTSM：神经生理学引导三流选择性状态空间模型——EEG-EOG实时疲劳检测8倍加速

论文信息

项目	内容
标题	NTSM: Neurophysiology-Guided Tri-Stream Selective State Space Model for Real-Time Driver Fatigue Detection with EEG-EOG Signals
PubMed	https://pubmed.ncbi.nlm.nih.gov/42647705/
发表	2026年8月
核心方法	三流选择性状态空间模型（Tri-Stream SSM）
输入	EEG + EOG
数据集	SEED-VIG + MPDDF（跨数据集验证）
加速	8.19× over Transformer E2CF

核心创新

1. 三流架构：EEG频谱流+EEG空间流+EOG时序流，各流独立SSM
2. 神经生理学先验：顶枕区激活=清醒，额颞慢波=疲劳
3. 选择性状态空间模型：Mamba风格SSM替代Transformer，8.19×加速
4. 跨数据集泛化：SEED-VIG训练→MPDDF测试92.95%准确率

方法详解

import torch
import torch.nn as nn

class SelectiveSSM(nn.Module):
    """选择性状态空间模型（Mamba风格简化版）"""
    
    def __init__(self, d_model: int, d_state: int = 16):
        super().__init__()
        self.d_model = d_model
        self.d_state = d_state
        
        # 输入投影
        self.proj_in = nn.Linear(d_model, d_model * 2)
        
        # SSM参数（可学习A, B, C, Δ）
        self.A = nn.Parameter(torch.randn(d_state, d_model) * 0.01)
        self.B = nn.Parameter(torch.randn(d_state, d_model) * 0.01)
        self.C = nn.Parameter(torch.randn(d_model, d_state) * 0.01)
        self.dt = nn.Parameter(torch.ones(d_model) * 0.1)
        
        self.proj_out = nn.Linear(d_model, d_model)
    
    def forward(self, x):
        """x: [B, T, D]"""
        x = self.proj_in(x)
        a, b = x.chunk(2, dim=-1)
        x = a * torch.sigmoid(b)  # Selective gate
        
        # 离散化SSM
        dt = torch.softplus(self.dt)
        A_bar = torch.exp(-dt.unsqueeze(0) * self.A)  # [D_state, D_model]
        
        # 简化前向扫描
        h = torch.zeros(x.shape[0], self.d_state, self.d_model, device=x.device)
        outputs = []
        for t in range(x.shape[1]):
            B_t = self.B * x[:, t:t+1, :].mean(dim=1, keepdim=True).transpose(0, 1)
            h = A_bar * h + B_t
            y = torch.einsum('bd,ds->bs', h.mean(dim=1), self.C)
            outputs.append(y.unsqueeze(1))
        
        out = torch.cat(outputs, dim=1)
        return self.proj_out(out + x)


class TriStreamNTSM(nn.Module):
    """NTSM: 三流选择性状态空间模型"""
    
    def __init__(self, n_classes=3):
        super().__init__()
        # 流1: EEG频谱
        self.spectral_ssm = SelectiveSSM(64)
        self.spectral_proj = nn.Linear(5, 64)  # 5频段→64
        
        # 流2: EEG空间
        self.spatial_ssm = SelectiveSSM(64)
        self.spatial_proj = nn.Linear(17, 64)  # 17通道→64
        
        # 流3: EOG时序
        self.eog_ssm = SelectiveSSM(32)
        self.eog_proj = nn.Linear(6, 32)  # 6个EOG特征
        
        # 融合+分类
        self.fusion = nn.Sequential(
            nn.Linear(64 + 64 + 32, 128),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(128, n_classes)
        )
    
    def forward(self, eeg_spectral, eeg_spatial, eog_features):
        # 三流各自SSM
        s = self.spectral_proj(eeg_spectral)  # [B, T, 64]
        s = self.spectral_ssm(s)
        
        p = self.spatial_proj(eeg_spatial)
        p = self.spatial_ssm(p)
        
        e = self.eog_proj(eog_features)
        e = self.eog_ssm(e)
        
        # 池化
        s = s.mean(dim=1)
        p = p.mean(dim=1)
        e = e.mean(dim=1)
        
        # 融合
        fused = torch.cat([s, p, e], dim=-1)
        return self.fusion(fused)


# 测试
if __name__ == "__main__":
    model = TriStreamNTSM(n_classes=3)
    
    # 模拟输入
    spec = torch.randn(4, 100, 5)   # 5频段×100时间步
    spat = torch.randn(4, 100, 17) # 17通道×100时间步
    eog = torch.randn(4, 100, 6)   # 6 EOG特征
    
    out = model(spec, spat, eog)
    print(f"NTSM输出: {out.shape}")
    print(f"参数: {sum(p.numel() for p in model.parameters()):,}")

实验结果

指标	NTSM	E2CF (Transformer)	提升
COR (Pearson)	0.8738	0.8321	+5.0%
RMSE	0.1294	0.1452	-10.9%
推理速度	8.19×	1×	8.19×
跨数据集准确率	92.95%	85.2%	+7.75%

神经生理学验证

脑区	清醒	疲劳	一致性
顶枕区	高激活	低激活	✅
额叶	β主导	θ/δ主导	✅
颞叶	正常	慢波增加	✅

IMS开发启示

价值	描述
8.19×加速	SSM替代Transformer，边缘部署友好
三流解耦	频谱+空间+EOG独立建模，可解释
跨数据集泛化92.95%	SEED-VIG→MPDDF验证
神经先验	模型决策与脑科学一致

总结

1. 三流SSM 8.19×加速：Mamba风格选择性SSM替代Transformer
2. 跨数据集92.95%：SEED-VIG→MPDDF泛化验证
3. 神经生理学一致：顶枕激活=清醒，额颞慢波=疲劳
4. 与DeltaGateNet互补：NTSM云端精细分析+DeltaGateNet边缘实时