Department of Internal Medicine, Kasturba Medical College, Manipal University, Mangaluru, Karnataka, India.
School of Information Sciences, Manipal Institute of Technology, Manipal University, Manipal, Karnataka, India.
J Healthc Eng. 2017;2017:5707162. doi: 10.1155/2017/5707162. Epub 2017 Nov 22.
Diagnosis of undifferentiated fever is a major challenging task to the physician which often remains undiagnosed and delays the treatment. The aim of the study was to record and analyze a 24-hour continuous tympanic temperature and evaluate its utility in the diagnosis of undifferentiated fevers. This was an observational study conducted in the Kasturba Medical College and Hospitals, Mangaluru, India. A total of ninety-six ( = 96) patients were presented with undifferentiated fever. Their tympanic temperature was recorded continuously for 24 hours. Temperature data were preprocessed and various signal characteristic features were extracted and trained in classification machine learning algorithms using MATLAB software. The quadratic support vector machine algorithm yielded an overall accuracy of 71.9% in differentiating the fevers into four major categories, namely, tuberculosis, intracellular bacterial infections, dengue fever, and noninfectious diseases. The area under ROC curve for tuberculosis, intracellular bacterial infections, dengue fever, and noninfectious diseases was found to be 0.961, 0.801, 0.815, and 0.818, respectively. Good agreement was observed [kappa = 0.618 ( < 0.001, 95% CI (0.498-0.737))] between the actual diagnosis of cases and the quadratic support vector machine learning algorithm. The 24-hour continuous tympanic temperature recording with supervised machine learning algorithm appears to be a promising noninvasive and reliable diagnostic tool.
未分化发热的诊断对医生来说是一项重大挑战，往往无法明确诊断，导致治疗延误。本研究旨在记录和分析 24 小时连续鼓膜温度，并评估其在未分化发热诊断中的效用。这是在印度芒格洛尔的卡斯特巴医疗学院和医院进行的一项观察性研究。共有 96 名（=96）表现为不明原因发热的患者。连续 24 小时记录其鼓膜温度。使用 MATLAB 软件对温度数据进行预处理，并从各种信号特征中提取特征，然后在分类机器学习算法中进行训练。二次支持向量机算法在将发热分为结核、细胞内细菌感染、登革热和非传染性疾病四大类方面的总体准确率为 71.9%。结核、细胞内细菌感染、登革热和非传染性疾病的 ROC 曲线下面积分别为 0.961、0.801、0.815 和 0.818。实际病例诊断与二次支持向量机学习算法之间存在良好的一致性[kappa=0.618（<0.001，95%置信区间（0.498-0.737））]。有监督机器学习算法的 24 小时连续鼓膜温度记录似乎是一种有前途的非侵入性和可靠的诊断工具。
