Page 68 - 南京医科大学自然版
P. 68
第46卷第7期
·1014 · 南 京 医 科 大 学 学 报 2026年7月
测能力有限等,这构成了现阶段的“数据挑战” 。 理解抑郁症脑网络异常的重要基础。常见的组合
[2]
随着神经影像、电生理、生物标志物及行为监测等 方式包括:sMRI 联合静息态 fMRI,用于从灰质形态
技术的发展,在同一患者身上获取多种客观指标成 和静息态网络同步性双层面表征脑网络失衡;sMRI
为可能,即形成多模态数据。在本综述中,多模态 联合 DTI,用于同时评估皮层萎缩与白质纤维束完
主要指以神经影像学数据为核心,同时包括 EEG、 整性;sMRI/DTI 联合任务态 fMRI 或磁共振波谱,从
语音与行为、临床与心理量表以及多组学等多源信 结构⁃功能⁃代谢耦合的角度刻画情绪和认知环路的
息。相比单一数据模态,多模态数据为抑郁症的临 异常。例如在一项研究中,部分患者表现为前额叶
床诊断、精准分型及病程动态评估提供了更立体全 皮层、海马等关键区域萎缩且功能连接减弱,呈现
面的研究视角,这也使得从脑结构特征、脑功能活 结构损害伴功能降低的特征;另一些患者结构指标
动、身心状态及日常行为表现等多个维度系统且综 相对正常,但 DMN 内部或 DMN 与认知控制网络间
合地认识抑郁症成为可能(表1)。 存在过度同步或解耦现象,以脑功能失衡为主要表
2.1 多模态神经影像的结构⁃功能⁃连接一体化表征 现 [24] 。这类多模态影像观察有助于将抑郁症从单
在神经影像领域,多种成像技术的联合应用是 一维度的有无异常转变为多维度的结构⁃功能表型,
表1 抑郁症多模态数据类型及其在机器学习诊断中的应用概述
Table 1 Overview of multimodal data types for depression and their application in machine learning diagnosis
Typical machine Representative
Data modality Key information reflected Main advantages and limitations
learning tasks literature
Neuroimaging Gray matter volume,cortical Classification:diagnosis(MDD Pros:high spatial resolution;maps 5,9-12,14,
thickness,white matter hyper⁃ vs. HC or others)& geriatric network topology;potential for objec⁃ 26-27
intensity burden;fractio ⁃ nal differential diagnosis; tive biomarkers;
anisotropy,mean diffusivity,Precision medicine:subtyping Cons:high cost;significant batch ef⁃
white matter tractography,& treatment response predic⁃ fects;overfitting risk in small sam⁃
and network efficiency tion ples
EEG & electro⁃ Temporal dynamics of brain Assessment:automated scree⁃ Pros:high temporal resolution;cost⁃ 15-19,28,
physiological activity,oscillatory rhythms,ning & severity grading; effective;portable;ideal for longitu⁃ 31,37
signals and neural synchronization Prediction:forecasting outco ⁃ dinal monitoring;
mes of drugs or stimulation Cons:low spatial resolution;sensi⁃
tive to artifacts;lacks standardization
Speech & digi⁃ Paralinguistic acoustic fea⁃ Detection:objective recogni⁃ Pros:non ⁃ invasive;high ecological 20-21
tal behavioral tures; ecological behaviors tion of depressive states; validity;enables passive,continuous
data such as physical activity lev⁃ Monitoring:continuous track⁃ sensing;
els and sleep rhythms ing of symptom fluctuation & Cons:Environmental confounders;
relapse prediction privacy risks;ground ⁃ truth labeling
difficulties
Clinical or psy⁃ Symptom dimensions,course Risk stratification:predicting Pros:low cost;scalable to multi⁃cen⁃ 2,3,8,23-25
chometric characteristics, comorbidity onset and relapse; ter cohorts;broad clinical coverage;
Scales & broad profiles,and treatmenthistory. Trajectory modeling:identify⁃ Cons:sbjective reporting bias;total
clinical data ing treatment⁃resistant or high⁃ scores mask symptom heterogeneity;
risk subgroups label noise
Multi⁃omics & Genetic susceptibility,tran⁃ Profiling:biological risk iden⁃ Pros:mechanistic insight;facilitates 22-24
fluid biomark⁃ scriptional and metabolic sta⁃ tification & subtyping; novel target discovery;
ers tes;inflammatory and endo⁃ Biotyping:linking molecular Cons:high cost;high dimensionality;
crine levels signatures to clinical pheno⁃ reproducibility challenges;requires
types large samples & rigorous QC

