基于声学流形解构与双轨风险拓扑的帕金森病无监督计算表型模型

Unsupervised Computational Phenotyping Model for Parkinson’s Disease via Acoustic Manifold Deconstruction and Dual-Track Risk Topology

摘要: 帕金森病(PD)早期起病隐匿且常伴随多系统退行性症状交织，极易导致临床误诊。基于数字语音信号的计算表型分析为PD的无创辅助诊断提供了新范式，但当前研究普遍受制于超高维特征冗余以及精细化亚型标签缺失两大数理瓶颈。为破解上述挑战，本文提出了一套融合高维拓扑流形解构与多尺度风险映射的智能诊断中枢。首先，针对高达754维的超高维声学特征，本文摒弃了常规的扁平特征筛选范式，设计了基于集成决策空间流形异质性散度(Heterogeneity Divergence)衰减的自适应降维算子。通过精细化量化决策树路径中的非线性纯度增益，成功解构了高维声学特征的“重尾稀疏”分布规律，精准锁定了以梅尔频率倒谱系数(MFCC)簇与可调Q因子小波变换(TQWT)多尺度阵列为核心的关键病理标志物子集，有效消解了维数灾难。其次，面对真实医疗场景中缺乏显式亚型“金标准”的困境，本文结合发声解剖学先验机理，将跨域异构特征重构为涵盖“运动控制、神经调控、混合结构”的三层六维数字化代理空间。在此基础上，本文首创“核心风险 + 结构辅助修正”的自适应双轨风险评分公式，并引入样本内相对百分位(Percentile)非线性映射平面。在零标签支持的严苛边界下，实现了对运动型风险、非运动型风险、混合型与不典型型四类精细化临床亚型的无监督精准聚类分型。大样本独立测试表明，本文构建的分型拓扑不仅在微观特征上具备极强的物理判别独立性，其宏观分类在后验验证集与独立测试集上分别取得了86.7%与81.7%的宏平均F1分数，展现出卓越的复现稳定性。本研究不仅为PD的高维声学数据降维提供了坚实的数理依据，更为复杂并发病症的精细化量化诊断提供了一套极具生物学解释性的辅助决策工具。

Abstract: The insidious onset and complex intertwining of multi-system degenerative symptoms in early Parkinson’s disease (PD) frequently lead to clinical misdiagnosis. Computational phenotyping based on digital speech signals offers a novel paradigm for non-invasive computer-aided diagnosis of PD. However, current research is severely constrained by two major mathematical bottlenecks: the curse of dimensionality caused by ultra-high-dimensional features and the lack of explicit labels for fine-grained clinical subtypes. To address these challenges, this paper proposes an intelligent diagnostic hub integrating high-dimensional topological manifold deconstruction and multi-scale risk mapping. First, to tackle the 754-dimensional acoustic features, we discard the conventional flat feature screening paradigm and design an adaptive dimensionality reduction operator based on the decay of manifold heterogeneity divergence within ensemble decision spaces. By accurately quantifying the nonlinear purity gain along decision paths, this mechanism successfully deconstructs the “heavy-tailed sparse” distribution of high-dimensional acoustic features. It accurately isolates a core subset of pathological biomarkers dominated by Mel-Frequency Cepstral Coefficient (MFCC) clusters and Tunable Q-factor Wavelet Transform (TQWT) multi-scale arrays, effectively resolving the curse of dimensionality. Second, to overcome the lack of gold-standard subtype labels in real-world medical scenarios, we reconstruct heterogeneous cross-domain features into a three-tier, six-dimensional digital proxy space—encompassing “motor control, neural regulation, and mixed acoustic structures”—guided by prior vocal anatomical mechanisms. Building upon this, we pioneer an adaptive dual-track risk scoring formulation that integrates “core risks and structural auxiliary corrections” and introduces an intra-sample relative percentile nonlinear mapping plane. Under the strict zero-label boundary condition, this framework achieves precise unsupervised clustering of four fine-grained clinical subtypes: motor risk, non-motor risk, mixed, and atypical. Large-sample independent testing demonstrates that our subtyping topology not only maintains strong physical discriminative independence at the micro-feature level but also achieves excellent reproducible stability, yielding macro-average F1 scores of 86.7% and 81.7% on the posterior validation and independent test sets, respectively. This research provides a solid mathematical foundation for the dimensionality reduction of high-dimensional acoustic data in PD and offers a highly biologically interpretable decision-support tool for the quantitative diagnosis of complex concurrent symptoms.