| 水利利,刘晨晨,李玉敏.基于脑电图参数的深度学习模型在首发精神分裂症患者疾病诊断及脑电图异常分级中的应用[J].四川精神卫生杂志,2025,(4):308-314.Shui Lili,Liu Chenchen,Li Yumin,A deep learning model for the diagnosis of first-episode schizophrenia and grading of EEG abnormalities using EEG signals[J].SICHUAN MENTAL HEALTH,2025,(4):308-314 |
| 基于脑电图参数的深度学习模型在首发精神分裂症患者疾病诊断及脑电图异常分级中的应用 |
| A deep learning model for the diagnosis of first-episode schizophrenia and grading of EEG abnormalities using EEG signals |
| 投稿时间：2024-07-16 |
| DOI：10.11886/scjsws20240716001 |
| 中文关键词: 深度学习 脑电图 精神分裂症 诊断 异常程度分级 长短期记忆模型 |
| 英文关键词:Deep learning model EEG Schizophrenia Diagnosis Abnormality grading Long short-term memory model |
| 基金项目: |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 背景 精神分裂症临床分型较多，异质性较大。以深度学习技术为代表的计算机技术为基于脑电图（EEG）的精神分裂症诊治和机制研究提供了巨大帮助，但基于我国人群的精神分裂症的相关研究仍相对缺乏。目的 探索基于脑电参数的深度学习模型在首发精神分裂症患者诊断及EEG异常分级中的应用，为提升精神分裂症的临床诊治能力提供参考。方法 选择2020年1月—2023年1月在阜阳市第三人民医院就诊的、符合《国际疾病分类（第10版）》（ICD-10）诊断标准的130例首发精神分裂症患者以及在本院进行体检的150名健康志愿者为研究对象。研究对象均接受EEG检查。基于EEG检查结果，建立长短期记忆（LSTM）深度学习网络模型，以十折交叉验证法进行机器学习研究，使用数据集中的90%作为建模组，10%作为验证组。以准确度、召回率、精确度和F1值、疾病诊断和EEG异常程度评估用时作为评价指标，并将深度学习模型表现与高年资医师评估结果进行比较。结果 LSTM诊断模型在建模组的精确度为（94.40±3.03）%，召回率为（94.30±3.23）%，准确度为（94.60±2.22）%，F1值为（94.20±2.20）%；在验证组中，精确度为（90.90±2.85）%，召回率为（92.20±1.14）%，准确度为（92.20±1.69）%，F1值为（91.50±1.78）%。LSTM诊断模型与高年资医师诊断精神分裂症的精确度、召回率、准确度以及F1值比较，差异均无统计学意义（χ2=1.500、0.750、2.722、1.056，P均>0.05）。LSTM模型评估EEG异常程度，在建模组中EEG异常程度评估准确度为（91.71±1.73）%，级别一精确度为（96.40±2.39）%，召回率为（94.77±1.40）%，F1值为（95.55±1.14）%，级别二精确度为（85.89±2.04）%，召回率为（88.10±6.18）%，F1值为（87.06±3.12）%，级别三精确度为（79.61±7.33）%，召回率为（81.79±9.87）%，F1值为（80.41±6.79）%。验证组EEG异常程度评估准确度为（85.61±6.16）%，级别一精确度为（91.43±6.25）%，召回率为（92.64±9.65）%，F1值为（91.56±4.83）%，级别二精确度为（71.17±19.02）%，召回率为（77.64±17.24）%，F1值为（71.88±11.33）%，级别三精确度为（90.00±21.08）%，召回率为（80.00±25.82）%，F1值为（81.67±19.95）%，LSTM模型与高年资医师评估精神分裂症EEG异常程度的精确度、召回率、准确度以及F1值比较，差异无统计学意义（χ2=0.098、0.036、0.020、0.336，P均>0.05）。LSTM模型诊断精神分裂症和EEG异常程度用时均少于高年资医师，差异均有统计学意义（t=57.147、43.104，P均<0.01）。结论 基于EEG的LSTM深度学习模型用于首发精神分裂症诊断和EEG异常程度分级的表现可能与高年资医师相当，且耗时较短。 |
| 英文摘要: |
| Background Schizophrenia is a highly heterogeneous disease with different clinical subtypes. Artificial intelligence technology represented by deep learning models has provided considerable benefits for the electroencephalogram （EEG）-based schizophrenia diagnosis， treatment and research， however， to date little research has been conducted regarding any of these benefits among Chinese schizophrenic patients.Objective To investigate the application of deep learning techniques utilizing EEG parameters for the diagnosis of first-episode schizophrenia and grading of EEG abnormalities in patients， with the aim of contributing to improved clinical diagnosis and treatment strategies for the disorder.Methods From January 2020 to January 2023， a total of 130 patients with first-episode schizophrenia who met the diagnostic criteria of International Classification of Diseases， tenth edition （ICD-10）， and attended at the Third People's Hospital of Fuyang， along with 150 health checkup examinees， were enrolled. All of them underwent EEG examination. An optimized long short-term memory （LSTM） deep learning model was developed utilizing EEG signals. Ten-fold cross-validation method was employed to evaluate the model's performance. The dataset was then split into two components： a training set （90%） for LSTM model development and a test set （10%） for validation. The accuracy， recall rate， precision， F1-score， schizophrenia diagnosis and EEG abnormality grading were used as evaluation indicators， and the results of the proposed model were compared to the assessments made by experienced psychiatrists.Results For schizophrenia diagnosis， the modeling group achieved the following performance metrics： precision （94.40±3.03）%， recall rate （94.30±3.23）%， accuracy （94.60±2.22）%， and F1-score （94.20±2.20）%. In the validation group， the corresponding metrics were precision （90.90±2.85）%， recall rate （92.20±1.14）%， accuracy （92.20±1.69）%， and F1-score （91.50±1.78）%. Statistical analysis revealed no significant differences between the LSTM diagnostic model and the experienced psychiatrists in terms of precision， recall rate， accuracy， and F1-score for schizophrenia diagnosis （χ2=1.500， 0.750， 2.722， 1.056， P>0.05）. The modeling group demonstrated an accuracy rate of （91.71±1.73）% in grading EEG abnormalities. For Grade 1 abnormalities， the modeling group reported a precision of （96.40±2.39）%， a recall rate of （94.77±1.40）%， and an F1-score of （95.55±1.14）%. In the case of Grade 2 abnormalities， the precision was （85.89±2.04）%， the recall rate was （88.10±6.18）%， and the F1-score was （87.06±3.12）%. For the more severe Grade 3 abnormalities， the modeling group's precision was （79.61±7.33）%， the recall rate was （81.79±9.87）%， and the F1-score was （80.41±6.79）%. Additionally， the validation group exhibited an accuracy rate of （85.61±6.16）%. The precision， recall rate， and F1-score for Grade 1 abnormalities were （91.43±6.25）%， （92.64±9.65）% and （91.56±4.83）%， respectively. For Grade 2 abnormalities， these metrics were （71.17±19.02）%， （77.64±17.24）% and （71.88±11.33）%. In the case of Grade 3 abnormalities， the precision was （90.00±21.08）%， the recall rate was （80.00±25.82）%， and the F1-score was （81.67±19.95）%. There was no significant difference in the accuracy， recall， accuracy and F1 value between LSTM model and senior doctors in evaluating the abnormal degree of EEG in schizophrenia （χ2=0.098， 0.036， 0.020， 0.336， P>0.05）. The LSTM model takes less time to diagnose schizophrenia and EEG abnormalities than senior doctors， and the differences were statistically significant （t=57.147， 43.104， P<0.01）.Conclusion The study utilizes an EEG-based LSTM deep learning model for diagnosing first-episode schizophrenia and grading EEG abnormalities， and the model not only matches the performance of experienced psychiatrists but also significantly reduces the time required for diagnosis. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |