Abstract
Although neuraxial techniques, such as spinal and epidural, are still considered as the gold standard for labor analgesia, there are some parturients who cannot receive neuraxial analgesia because of pre-existing conditions, or who request analgesia other than epidural block. An alternative analgesia is remifentanil, which is a relatively new, very potent and short-acting opioid. It has been shown to be effective in the relief of labor pain, but reports to date have failed to find the optimal dosing regimen. A challenge to a systemic opioid is that it must match the unique time course of labor pain. A continuous infusion is not ideal, as the parturient experiences no pain between contractions. Moreover, a continuous infusion during times in which the patient does not experience pain, may increase the risks of respiratory depression, sedation and nausea. The continuous infusion also increases the amount of the drug to which the fetus is exposed. Designing an optimal dosing regimen necessitates the prediction of the pace of contractions, so that the drug can be given shortly before the pain of the contraction begins. The prediction and thus drug administration should be made early enough to allow for the administration of intravenous analgesia that will have maximal efficacy during contractions, little effect between contractions, and minimal impact on the fetus. Towards such a need, we propose a knowledge-assisted sequential pattern analysis framework to predict the changes in intrauterine pressure, which indicate the occurrence of labor contractions. The proposed framework predicts in real time and provides a prediction multiple seconds before a contraction occurs, so as to assist in designing optimal administration strategies of remifentanil in labor. The proposed framework first selects a group of patients, from the stored record, who share similar demographic and obstetrical information with the current patient of interest. Second, it develops a sequential association rule mining approach to learn the patterns of the contractions from the historical patient tracings of the selected patients. Third, a sequential association rule-based collaborative filtering strategy is designed to dynamically select a training dataset from the historical patient tracings, as well as from the most recent training time series of the patient of interest. The training set is used for training a set of prediction models. A k-nearest neighbors (k-NN) based least squares support vector machine (LS-SVM) approach with heuristic parameter tuning is proposed to conduct the long-term time series prediction. A post-prediction process is also incorporated to further enhance the prediction results. Because to the best of our knowledge, there has been no previous study to predict future contractions, this work can be considered as a pioneer in the field. We evaluate the performance of the proposed framework using actual data from anonymous patients with varied contraction patterns. The data include patient demographic and obstetrical information, and measured intrauterine pressure time series. Overall, the proposed framework outperforms several well-known prediction methods, and it accomplishes that in real time. Meanwhile, experiments that compare each component with some other famous algorithms are conducted. The promising experimental results show that all proposed components improve the prediction precision, and the proposed framework achieves the effectiveness, robustness and efficiency that are needed for designing the optimal dosing regimen of remifentanil.