Prediction Of Uterine Contractions Using Knowledge .

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.2the next contraction multiple seconds before it occurs, along-term prediction is required. The ever changing laborcontraction pattern and diversity across different patients haveposed many challenges to a predication model with regard todenoising, adaptive long-term time series prediction, knowledge discovery from the patients’ tracings, and personalizingthe prediction.To develop an accurate model for predicting contractionsto allow for maximal analgesic efficacy, we design a sequential association rule-based collaborative training datasetselection component to dynamically select a training datasetfrom historical patient tracings (HT ) and the current patient’smost recent training time series. A k-nearest neighbors (kNN) based least squares support vector machine (LS-SVM)approach with heuristic parameter tuning is then employed toconduct long-term time series prediction. A post-predictionprocess is also proposed to further enhance the predictionresults. The framework conducts the prediction in real time.To the best of our knowledge, this is the first study thataddresses the potential application of a sequential patternanalysis to the practice of delivering analgesia to laboringwomen. Engineers and obstetric-anesthesiologists are workingtogether to pinpoint the exact time for delivery of ultrafastacting opioids (remifentanil) to optimize pain management andminimize dangers. While our research primarily addresses themethodology for developing the sequential pattern analysis,the driving force behind this approach is the health, safety,and well-being of the mother and the infant.The rest of the paper is organized as follows. Section IIbriefly discusses some related studies. The structure of theproposed framework and the technical details are presented inSection III. Experimental evaluation, results, and discussionsare in Section IV. Section V gives the conclusion of the paper.Association rule mining is a significant and recognized technique of data mining [17]–[19]. Its goal is to extract interestingcorrelations, frequent patterns, associations or causal structuresamong sets of items in the transaction databases or otherdata repositories. Association rule mining discovers interestingassociation rules that satisfy the predefined minimum supportand confidence from a given database. Its primary applicationis market basket analysis. Interesting association rules can beused as the basis for planning marketing activities. In addition,association rules are employed in other application areasincluding web usage mining [20], intrusion detection [21]and bioinformatics [22]. For example, a collaborative filteringmethod based on the association rule mining technique wasdeveloped to efficiently generate a list of webpage recommendations [20].In our proposed framework, collaborative filtering is usedto solve the contraction prediction problem. If the patient’scontraction pattern matches with the contraction patterns ofselect previous patients, it is likely that the current patient’snext contraction will be similar to, if not the same as, thenext contraction of other patients in the patient database.Based on this assumption, our framework builds a databaseof contraction patterns (in the form of sequential associationrules) from the historical patients’ labor tracings and searchesfor the matching sequential association rules of the currentpatient’s contraction pattern to predict the pattern of the currentpatient’s next contraction. The idea is that if the currentpatient’s contraction pattern matches with the condition partof a rule (i.e., the contraction pattern of the past patients), thepatient’s next contraction pattern is predicted to be the same asthe consequent part of the rule (i.e., the next contraction patternof previous patients). Through the matching process, we canselectively utilize the available data to train the predictionmodels. The technical details are provided in Section III.II. R ELATED W ORKCollaborative filtering (CF) is the process of filtering forinformation using techniques involving collaboration amongmultiple data sources [10]. The rationale behind collaborativefiltering is that a user’s preferences, interests, or behavior canbe predicted by leveraging the preferences, interests, or behavior of the communities of similar users. Collaborative filteringis a successful approach to build recommender systems [11],which assist people in decision making on merchandise, webpages, movies, music, and restaurants. The developers of oneof the first recommender systems, Tapestry [12], started to usethe phrase ‘collaborative filtering’ (CF). The early generationof collaborative filtering systems, such as GroupLens [13], usethe user ratings to calculate the similarity or weight betweenusers or items and make predictions or recommendations.Lately, the memory-based CF methods are deployed intocommercial systems such as Amazon and Barnes and Noble. Itis due to the fact that the memory-based CF methods are easyto implement and highly effective [14], [15]. The CF systemspromote a greater customer loyalty, higher sales, and moreadvertising revenues [16]. However, CF is primarily linked toe-commerce and has not yet been applied to healthcare relatedstudies.III. P ROPOSED F RAMEWORKThe intrauterine pressure signal is measured by the intrauterine pressure catheter. Usually, an intrauterine pressuretracing is composed of intermittent peaks one after anotherwith rest periods in between. The start of a peak correspondsto the start of a contraction. The time from the beginningof one contraction to the beginning of the next contractionis identified as ‘period’. Fig. 1 provides an example of theintrauterine pressure time series. In this example, the periodis approximately 5 minutes.Fig. 1.Intrauterine Pressure Time SeriesThe original intrauterine pressure time series are sampledat each quarter second within the value range [0,100]. DueCopyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.3to some system constraints, patients’ movements, or otherunforeseen interference, the intrauterine pressure time seriesare often contaminated by noise and may suffer from loss ofdata. Poor data quality significantly influences the performanceof the prediction. Therefore, a preprocessing step is necessaryto reduce the noise as much as possible. The pattern ofthe noise in intrauterine contraction time series varies fromtime to time, and it is difficult to filter out the noise by aregular low pass filter. We subsample the data at each second,and normalize the values to [0,100]. In addition, we furthersubsample the data and detect the peak points, which arevery crucial to determine the curve shape. Based on theseselected points, the shape-preserving piecewise cubic Hermiteinterpolation method is applied to rebuild the signal.Fig. 2 shows the flowchart of the proposed predictionframework. The inputs to the framework include a trainingtime series d, HT , and a testing time series td. Both d andtd are derived from the intrauterine pressure time series ofthe current patient of interest. HT contains multiple denoisedpatients’ intrauterine pressure time series. The output of theframework is intrauterine pressure values predicted multipleseconds ahead. The framework mainly contains four parts according to the functionalities: 1) collaborative training datasetselection, 2) heuristic parameter tuning for LS-SVM, 3) k-NNbased LS-SVM for long-term time series prediction, and 4) thepost-prediction process, which includes boundary constraint,multi-value integration and vertical correction components.A heuristic strategy is proposed to tune the parameters forLS-SVM with radial basis function (RBF) kernel, includingthe regularization factor and the Gaussian kernel parameter.This strategy will be detailed in another paper. The initialvalue for the regularization factor is set based on the varianceof the output in the training dataset, and the variance of thenoise. Meanwhile, modeling error is taken into considerationto adjust the value of regularization factor. The tuning processconverges quickly because of taking the error feedback. Inaddition, the searching range of the Gaussian kernel parameteris decided based on the distance of the input vectors of thetraining instances. If real-time prediction was not required, thebest option would be using the output of collaborative trainingdataset selection as the input for parameter tuning. However,real-time prediction is required in this specific application.In order to make it more efficient, parameter tuning andcollaborative training dataset selection are executed in parallel.Meanwhile, the selected contractions from HT share similarpatterns with the contractions in d, so it is reasonable touse d to determine the parameters. The parameter tuningprocess guarantees that the proposed framework is able toreach its best performance. k-NN based LS-SVM approachdescribed in [23] is employed here for long-term time seriesprediction. In this approach, the k-NN component selects a setof instances from the training dataset for each of the testinginstance to train the LS-SVM model. The strategy reduces thecomplexity of training the LS-SVM models and also improvesthe prediction accuracy. We present the details of the novelcollaborative training dataset selection strategy and the postprediction process in the following subsections.Fig. 2.Flowchart of the Proposed FrameworkA. Collaborative Training Dataset SelectionCollaborative filtering (CF) is a technique that uses theknown preferences of a group of users to make recommendations or predictions of the unknown preferences for otherusers [11]. In our framework, a CF component is proposedto use the known sequential contraction patterns of a groupof past patients to assist in predicting the unknown comingcontraction for other patients. Fig. 3 shows the processes of thecollaborative training dataset selection component (enclosedby the dashed lines). It is a rule matching based datasetselection. The rule matching process takes inputs from twobranches as shown in Fig. 3.Fig. 3.Collaborative Training Dataset Selection ComponentNext, we introduce the designed discretization method,propose a sequential association rule mining approach formining the frequent sequential contraction patterns, and detailthe process of the rule matching based collaborative trainingdataset selection.1) Contraction Feature Extraction & Discretization: Thecontraction pattern is determined by the combination of boththe height and the period of a contraction. These two featuresextracted for contractions are numerical, while the sequentialassociation rule mining algorithm is to discover interestingsequential relations between categorical variables in largedatabases. To analyze the sequential association relationshipsamong contractions, it is necessary to discretize the features.We adopt the equal-width discretization method in ourframework considering that the discretization method shouldCopyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.4facilitate the association rule mining process. On one hand,there should be a certain number of rules generated. On theother hand, the generated rules should be meaningful andreasonable. Even though the equal-width discretization methodis relatively straightforward, it suits our application better thanother discretization methods. Table I and Table II show themapping tables used in our experiments, which are decidedbased on domain knowledge. Contractions detected from boththe training time series d and HT are discretized using thesame mapping tables.TABLE ID ISCRETIZATION M APPING TABLE FOR H EIGHT (in mmHg)Height (ht) rangeht 4040 ht 5555 ht 7070 ht 8585 ht 100ht 100Nominal markerabcdefTABLE IID ISCRETIZATION M APPING TABLE FOR P ERIOD (in seconds)Period (pd) rangepd 2525 pd 8585 pd 145145 pd 205205 pd 265265 pd 325325 pd 385pd 385Nominal markerABCDEFGHFor the intrauterine pressure time series given in Fig. 1,the period is 5 minutes, i.e., 300 seconds. Consequently, theperiod is discretized as ‘F’. Let the height of the peak be 75mmHg, then it is discretized as ‘d’. Therefore, the contractionshown in Fig. 1 can be described as ‘dF’. For consecutivecontractions one after the other, they can be represented asa sequence of paired letters in the similar manner, i.e., adiscretized contraction features series.2) Sequential Association Rule Mining Algorithm: Wedevelop a sequential association rule mining algorithm todiscover interesting sequential patterns from the intrauterinepressure time series. The input of the sequential associationrule mining process is the discretized contraction feature seriesderived from HT . An item is one contraction and it containstwo letters, representing its height and period, respectively.A sequential itemset (wX yZ) describes the occurrence oftwo consecutive contractions, where w and y are discretizedheights, and X and Z are discretized periods. The process returns a set of rules (R) as the output in the form of wX yZ.HT contains multiple patient intrauterine pressure tracings,

A post-prediction process is also proposed to further enhance the prediction results. The framework conducts the prediction in real time. To the best of our knowledge, this is the first study that addresses the potential application of a sequenti