融合贝叶斯估计和主动学习的半隐马尔科夫模型及在动态心电分割上的应用

For the world's highest mortality cardiovascular disease, prevention has become a consensus. Wearable real-time ECG monitoring is undoubtedly one of the most effective means. In the dynamic long-term monitoring environment, the signal noise is serious and unpredictable, which leads to the essential difficulty for the feature extraction. Hidden Markov Model (HMM) has better robustness due to the integration of signal statistics. However, in the face of complex and variable wearable ECG signal, the lack of state duration distribution, the robustness of HMM is difficult to guarantee. In view of this, the project increases the underlying flexibility of the model by relaxing the Markov assumption to enable it to incorporate any form of duration distribution. The modified and extended Viterbi algorithm is added to overcome the implicit hypothesis. The joint logistic regression and naive Bayesian theory use the probability statistics information to determine the observed probability matrix of the model. Based on these two aspects, a new semi-hidden Markov model (HSMM) is developed. Considering the wearable Holter that is difficult to control with complex signal quality, waveform and rhythm, it is necessary to establish high cost of expert mark to establish a stable new HSMM model. This project builds a wearable dynamic ECG signal optimization mechanism based on integrating quality assessment system and signal active selection mechanism. And an enhanced HSMM algorithm is developed, which can improve the robustness of accurate segmentation for wearable dynamic ECG signals. This work could provide theoretical and technical support for medical clinical applications such as early screening and long-term monitoring of cardiovascular disease.

面对全球死亡率最高的心血管疾病，治防同步，以防为主已成共识，穿戴式实时心电监控无疑是最有效的手段之一。动态长时间监测环境下，信号复杂多变,导致特征提取遇到本质困难。隐马尔可夫模型（HMM）因融入信号统计特性而具有更好的鲁棒性，但面对复杂多变的穿戴式心电，缺少状态持续时间分布控制的HMM稳健性难以保证。鉴于此，本项目通过松弛马尔可夫假设增加模型底层灵活性使其能够纳入任意形式的持续时间分布，修正和扩展Viterbi算法，联合Logistic回归和贝叶斯理论发展新的半隐马尔可夫模型（HSMM）。为提高模型的泛化能力，本项目基于质量评估和主动学习算法构建穿戴式动态心电信号优选机制，强化状态持续分布密度函数和状态幅值分布函数的泛化能力，发展增强型HSMM算法，提高穿戴式动态心电信号精准分割稳健性，进而为心血管病早期筛查和长期监测等医学临床应用提供理论和技术支持。

穿戴式动态心电信号的特征波段的精准提取是穿戴式医疗设备应用于临床心血管病监测和早期预警中亟待解决的关键问题，本项目基于质量评估和主动学习发展增强型的隐马尔可夫算法，实现动态心电各个波段实时精准提取。研究内容：（1）建立了一个具有50085条记录的可穿戴式心电图质量等级数据库，其中包括高质量/中等质量/低质量三个质量等级；（2）发展泛化bSQI动态心电信号质量评估算法，研究多标志信息融合的质量评估模型，实现动态信号的轻量级快速质量评估；（3）提出一种基于三层小波散射网络和深度学习LSTM的质量优选方法，通过对信号的深入分析，可以提取出更系统、更全面的特征，实验结果和实际数据验证表明，该方法具有较高的准确性、鲁棒性和计算效率；（4）融合主动学习和集成回归方法，构建四种主动叠加模型，优选少量最有价值的未标记样本训练模型，实现心率的精准估计；（5）基于优选数据，松弛马尔可夫假设增加模型底层灵活性使其能够纳入波形持续时间的统计概率分布，提取ECG信号的四个特征向量作为隐马尔可夫模型的观察序列，修正和扩展Viterbi算法，发展双分布函数的增强型双向半隐马尔科夫模型，并开展实验验证。