Heda Vishrut, Dogra Saanvi, Kouznetsova Valentina L, Kumar Alex, Kesari Santosh, Tsigelny Igor F
Scholars Program, CureScience Institute, San Diego, CA 92121, USA.
MAP Program, University of California San Diego, La Jolla, CA 92093, USA.
Int J Mol Sci. 2025 Mar 4;26(5):2280. doi: 10.3390/ijms26052280.
Diagnostic practices for schizophrenia are unreliable due to the lack of a stable biomarker. However, machine learning holds promise in aiding in the diagnosis of schizophrenia and other neurological disorders. Dysregulated miRNAs were extracted from public sources. Datasets of miRNAs selected from the literature and random miRNAs with designated gene targets along with related pathways were assigned as descriptors of machine-learning models. These data were preprocessed and classified using WEKA and TensorFlow, and several classifiers were tested to train the model. The Sequential neural network developed by authors performed the best of the classifiers tested, achieving an accuracy of 94.32%. Naïve Bayes was the next best model, with an accuracy of 72.23%. MLP achieved an accuracy of 65.91%, followed by Hoeffding tree with an accuracy of 64.77%, Random tree with an accuracy of 63.64%, Random forest, which achieved an accuracy of 61.36%, and lastly ADABoostM1, which achieved an accuracy of 53.41%. The Sequential neural network and Naïve Bayes classifier were tested to validate the model as they achieved the highest accuracy. Naïve Bayes achieved a validation accuracy of 72.22%, whereas the sequential neural network achieved an accuracy of 88.88%. Our results demonstrate the practicality of machine learning in psychiatric diagnosis. Dysregulated miRNA combined with machine learning can serve as a diagnostic aid to physicians for schizophrenia and potentially other neurological disorders as well.
由于缺乏稳定的生物标志物，精神分裂症的诊断方法并不可靠。然而，机器学习有望帮助诊断精神分裂症和其他神经疾病。从公共来源提取了失调的微小RNA（miRNA）。从文献中选择的miRNA数据集以及具有指定基因靶点和相关通路的随机miRNA被指定为机器学习模型的描述符。这些数据使用WEKA和TensorFlow进行预处理和分类，并测试了几个分类器以训练模型。作者开发的顺序神经网络在测试的分类器中表现最佳，准确率达到94.32%。朴素贝叶斯是次优模型，准确率为72.23%。多层感知器（MLP）的准确率为65.91%，其次是霍夫丁树，准确率为64.77%，随机树的准确率为63.64%，随机森林的准确率为61.36%，最后是ADABoostM1，准确率为53.41%。由于顺序神经网络和朴素贝叶斯分类器的准确率最高，因此对其进行测试以验证模型。朴素贝叶斯的验证准确率为72.22%，而顺序神经网络的准确率为88.88%。我们的结果证明了机器学习在精神疾病诊断中的实用性。失调的miRNA与机器学习相结合可以作为医生诊断精神分裂症以及潜在的其他神经疾病的辅助手段。
