 Technical advance
 Open Access
 Open Peer Review
 Published:
A clustering approach for detecting implausible observation values in electronic health records data
BMC Medical Informatics and Decision Makingvolume 19, Article number: 142 (2019)
Abstract
Background
Identifying implausible clinical observations (e.g., laboratory test and vital sign values) in Electronic Health Record (EHR) data using rulebased procedures is challenging. Anomaly/outlier detection methods can be applied as an alternative algorithmic approach to flagging such implausible values in EHRs.
Methods
The primary objectives of this research were to develop and test an unsupervised clusteringbased anomaly/outlier detection approach for detecting implausible observations in EHR data as an alternative algorithmic solution to the existing procedures. Our approach is built upon two underlying hypotheses that, (i) when there are large number of observations, implausible records should be sparse, and therefore (ii) if these data are clustered properly, clusters with sparse populations should represent implausible observations. To test these hypotheses, we applied an unsupervised clustering algorithm to EHR observation data on 50 laboratory tests from Partners HealthCare. We tested different specifications of the clustering approach and computed confusion matrix indices against a set of silverstandard plausibility thresholds. We compared the results from the proposed approach with conventional anomaly detection (CAD) approaches, including standard deviation and Mahalanobis distance.
Results
We found that the clustering approach produced results with exceptional specificity and high sensitivity. Compared with the conventional anomaly detection approaches, our proposed clustering approach resulted in significantly smaller number of false positive cases.
Conclusion
Our contributions include (i) a clustering approach for identifying implausible EHR observations, (ii) evidence that implausible observations are sparse in EHR laboratory test results, (iii) a parallel implementation of the clustering approach on i2b2 star schema, and (3) a set of silverstandard plausibility thresholds for 50 laboratory tests that can be used in other studies for validation. The proposed algorithmic solution can augment human decisions to improve data quality. Therefore, a workflow is needed to complement the algorithm’s job and initiate necessary actions that need to be taken in order to improve the quality of data.
Background
Implausible observations in electronic health records
Data stored in Electronic Health Records (EHR) offer promising opportunities to advance healthcare research, delivery, and policy. Provision of these opportunities is contingent upon high quality data for secondary use. Data quality concerns, however, have hampered secondary use of EHR data [1, 2]. The increasing throughput of EHR data constantly deposited into clinical data research networks have cultivated new opportunities for utilizing innovative statistical learning methods to improve quality.
Plausibility is a dimension of data quality that represents whether EHR data values are “believable”. [3]. It is quite possible to witness an implausible observation (IO) in EHR data, such as a negative A1c value. An IO is extremely unlikely to signify a fact about a patient and may represent an underlying data quality issue. Detecting such IOs in EHR data are difficult for two reasons. First, gold standards are not always available for all clinical observations to set cutoff thresholds for an implausible observation. Second, even when gold standards are present, detection of outofrange observations requires limits to be determined for each type of observation, potentially customized at each institution (due to variance in normal ranges). This rulebased approach is becoming increasingly impractical given the everyincreasing diversity of observations, ontologies and measurement units across institutions.
With the abundance of unlabeled data (e.g., vital signs) in EHR repositories, unsupervised learning can offer solutions for characterizing clinical observations into meaningful subgroups. In unsupervised learning, the machine develops a formal framework to build representations of the input data to facilitate further prediction and/or decision making [4]. We focus on records of laboratory results and vital signs in the EHR. The variance and variety of these domains make it particularly difficult to manually assign rules for identifying implausible values. Conceptualizing implausible laboratory test results and vital sign values in EHRs as outliers, we propose and evaluate the feasibility of an unsupervised clustering approach for detecting implausible EHR vital sign and laboratory test values in large scale clinical data warehouses.
Outlier detection
Outlier detection has been widely applied in medical informatics for addressing different issues, such as detecting unusual patientmanagement actions in ICU [5], deriving workflow consensus from multiple clinical activity logs [6], characterizing critical conditions in patients undergoing cardiac surgery [7], discovering unusual patient management [8], alert firing within Clinical Decision Support Systems [9], finding clinical decision support malfunctions [10], identifying high performers in hypoglycemia safety in diabetic patients [11], and classifying the influence factor in diabetes symptoms [12].
In outlier detection, the objective is discriminate nonconforming observations from a larger group of observations that conform to similar attributes (inliers) [13]. Due to discrepancies in defining outlyingness, various outlier detection techniques have been proposed [14]. Both parametric (modelbased) and nonparametric (modelfree) approaches are common for detecting outliers. Nonparametric approaches do not assume apriori statistical models and therefore can be more suitable for clinical data that are often irregularly sampled. Distance and densitybased techniques utilize different proximity (distance) measures to identify outlying observations. Distancebased methods use local distance measures, such as the Mahalanobis distance, to identify outliers based on distance from the nearest neighbors [15, 16]. Although these methods are popular and often scale to large data [17], their performance decreases in high dimensional spaces [14]. Moreover, the performance of distancebased methods relies on a distance threshold for identifying outliers, and in our experience (including the results in this study), can lead to high false positive rates when applied to human biological data.
Densitybased techniques, in contrast, assume that the data density around a inlier data point is similar to the density around its neighbors. Clustering techniques can utilize a densitybased approach, e.g. to identify sparse clusters for outlier detection [18]. In fact, clustering techniques can combine attributes from both the distancebased and the densitybased methods. We will use a hybrid hierarchicalkmeans clustering method in this study, in order to perform a more meaningful detection of outlying observations from human biological data based on both distance and sparsity.
The hierarchical Kmeans clustering method
The goal in unsupervised clustering is to partition data points into clusters with small pairwise dissimilarities [19, 20]. Kmeans [21] is one of the most popular unsupervised clustering algorithms [19], for its simplicity and efficiency. It is a topdown algorithm that aims to minimize the distortion measure by iteratively assigning data points to cluster centroids to meet a convergence criterion [4, 19]. Kmeans is sensitive to outliers [22, 23]. Although this property is often considered a weakness, sensitivity to outliers makes Kmeans a good algorithm for our purpose of identifying rare events in clinical observations.
For a vector of observations x(1), x(2), …, x(n), where x(i) ∈ℝ^{n}, the Kmeans algorithm aims to predict k centroids and assign the data points to each centroid to form clusters C(i), while minimizing the average withincluster dissimilarity, as follows:

1
Randomly initialize k cluster centroids μ(1), …, μ(k)

2
For a given cluster assignment C, iterate the following steps until the cluster assignments do not change:

a.
$$ C(i):= \underset{k}{\mathrm{argmin}}\parallel x(i)\mu (k){\parallel}^2. $$

b.
$$ \mu (k):= \frac{\sum_{i=1}^nI\left\{C(i)=k\right\}x(i)}{\sum_{i=1}^nI\left\{C(i)=k\right\}}. $$

a.
This specification makes Kmeans’ performance dependent on initialization of a few hyperparameters, including the number of clusters and the initial cluster centroids [19, 20]. The dependency of Kmeans on random initialization of the cluster centers often results in the algorithm’s performance being unreliable, when the number of iterations are small [23]. In contrast to Kmeans, hierarchical clustering is a bottomup or agglomerative approach that does not require knowing the number of clusters in advance. As a result, its performance is not dependent on random initialization of k cluster centroids. However, compared with the Kmeans algorithm, hierarchical clustering algorithm is more computationally intensive.
Hybrid hierarchicalkmeans (HKmeans) clustering [23] combines the strengths of hierarchical clustering in initializing the cluster centroids, and improves efficiency of the Kmeans algorithm. The HKmeans algorithm relaxes the dependency of Kmeans algorithm on random initialization of cluster centroids by first computing hierarchical clustering, cutting the tree in k clusters, computing the centroids for each cluster, and then using the centroids as the initial cluster centers to run Kmeans. [23]. This hybrid approach accelerates the Kmeans procedure, and thereby, improves the overall learning [23].
Performance of the HKmeans algorithm still depends on approximation of the number of clusters, k. Initializing the number of clusters for the algorithm is a challenging problem, for which a number of adhoc (or intuitionbased) solutions are available [24,25,26]. Unfortunately, most of the available solutions do not scale up to large datasets.
Methods
Approach
We propose an unsupervised approach using the hierarchicalkmeans method to detect outlier values in lab tests and vital sign values based on two hypotheses. First, we hypothesize (hypothesis #1) that implausible observations – that are not due to systematic human error – must be sparse (infrequent) in datasets that contain large amounts of data. For example, we do not expect to frequently see a blood pressure record of 1090/80 in an EHR repository. Given ontological harmony, we can use data from EHRs to compare an individual data point with a large set of similar data points and identify implausible observations. Therefore, we also hypothesize (hypothesis #2) that a wellspecified unsupervised clustering algorithm should be able to partition clinical observations into meaningful clusters, from which we can extract clusters with sparse populations (very small members of data points) as implausible observations.
Figure 1 illustrates the rationale behind our second hypothesis. The bellshaped curve in Fig. 1 is a probability density function for a laboratory result in EHR – extracted and visually modified from Cholesterol in HDL. The hypothetical thresholds for normal ranges and implausible values are delineated on the plot. An unsupervised clustering algorithm partitions the lab values into n clusters, each of which embrace a number of data points. We can obtain the number of data points in each cluster and flag clusters with population smaller than a certain threshold as anomalies. Unlike conventional methods such as using standard deviation to identify anomalies, the clustering approach would provide a more flexible solution that is densitybased. In addition, the clustering approach should be able to detect implausible observations regardless of their values.
To test these hypotheses, we implemented the clustering solution on EHR observation data from a portion of the Research Patient Data Registry (RPDR) from Partners HealthCare [27], using a hybrid hierarchical Kmeans clustering algorithm. Specifically, we used kluster, an R package that uses iterative sampling to produce scalable and efficient cluster number approximation solutions in large datasets [28]. We set different ratios for flagging a cluster as anomalous. To compare the results against conventional methods, we also performed anomaly detection using conventional approaches (CAD) using standard deviation and Mahalanobis distance [29, 30] with different configurations for each.
We measured the performance for each algorithm against a set of silverstandard high and low implausible thresholds that we manually curated based on literature search and expert judgment, and validated using data distributions in RPDR – Table of silver standards is available in the Appendix (Table 4).
Data
The data used in this study contained ~ 720 million rows of medical record data representing 50 laboratory observations and a small set of common vital signs, all with distinct Logical Observation Identifiers Names and Codes (LOINC) codes. This data represented contained all results for these 55 observations in the RPDR for patients at Massachusetts General Hospital and Brigham and Women’s Hospital between 2001 and 2018. The RPDR derives these data from various clinical systems, such as the institutional Clinical Data Repository (CDR), the Longitudinal Medical Record (LMR) system, and EpicCare. The lab data are assigned LOINC codes by the Clinical Data Repository team. Otherwise, the data are not modified from the source system before being provided to researchers. On average, each observation type contained more than 14 million data points (range 45,000 to > 69 million). These data are made available for the Accessible Research Commons for Health (ARCH), a 12site PCORnet Clinical Data Research Network (CDRN).
Implementation
Due to the size of data for each type of lab or vital sign observations, we parallelized the implementation of our clustering solution for identifying the implausible lab observations. Through the parallelization, we also incorporate a hyperparameter that represent the sparsity assumption in the approach. We used R statistical language and high performance computing cluster – provided by the Partners HealthCare’s Enterprise Research Infrastructure & Services – to implement the algorithm, described in following steps (also Fig. 2):

1.
Extract data on observation x, stored as db(x), from RPDR.

2.
Shuffle db(x) randomly to avoid any specific sorting for parallelization.

3.
Break db(x) into j folds such that each fold has n (or fewer) data points.

We controlled the threshold for flagging clusters as implausible (\( \boldsymbol{\alpha} =\frac{\mathbf{1}}{\boldsymbol{n}} \)) through the selection of the number of data points for parallelization. We evaluated the performance of the proposed methodology using eight thresholds for α: 1/500, 1/1000, 1/2000, 1/3000, 1/4000, 1/5000, 1/6000, and 1/10,000.


4.
Begin parallel computing:

Extract the subset of db(x) for fold j, db(x j).

Scale db(x j) and transform the values to the 3rd power – to focus on distribution tails.

Apply kluster procedure [28]. to db(x j) to identify k clusters.

Compute HKmeans clustering and assign data points to clusters cluster C(1), …, C(k).

Count number of data points in each cluster C(k), p_{k}.

Flag all data points in C(k) as implausible, where p_{k} ≤ α.


5.
Produce a report containing all flagged rows.
To measure performance, we compute confusion matrix indices, including false/true positives, false/true negatives, sensitivity, specificity, and fallout across the eight α s. True positive in the confusion matrix represents the number of truly implausible observations (as identified from the silverstandard labels) that were also identified by the clustering approach as implausible.
We also evaluated whether our proposed clustering approach performed better than conventional anomaly detection (CAD) methods, namely identifying implausible values using standard deviation and anomaly detection with Mahalanobis distance (Table 1).
Results
Sensitivity and specificity metrics for all implementations can be seen in Tables 2 and 3. Here we present a summary, beginning with specificity (i.e., how many data points identified as plausible by the clustering approach are truly plausible).
Specificity
In all of the 50 lab observations, the clustering approach performed with a specificity greater than 0.9997 (Table 1). The best specificities were often obtained from the most stringent α (1/10,000), which identifies a cluster as implausible only if its population is 1/10,000 of the data subset. In this configuration, we used subsets of 10,000 data points for parallel computing. The 10,000 data points were partitioned into n clusters, and the cluster with 1 data point was identified as implausible. The lowest specificity, 0.9938, was from the most liberal configuration with α = 1/500. Overall, we found that specificity increases as α decreases (Fig. 3).
Sensitivity
Sensitivity focuses on true positives, and in our case, represents how many of the implausible observations were picked up by the algorithm. Our sensitivity results we less consistent than the specificities we obtained from the clustering approach. We did not have any implausible observations in 9 of the 50 lab tests – i.e., no true positives in 18% of the labs. In the 41 remaining labs, we obtained the best performance from the most liberal configuration α, where 1 in 500 data points was flagged as implausible (Table 2). It is important to evaluate the sensitivity results considering the sparsity of positives (implausible observations) in data. The number of implausible observations in the 41 labs ranged from 1 to over 39,000, representing an average of 0.0576% of the labs. Considering sparsity, a sensitivity over 0.85 would pick up most of the implausible observations. We obtained sensitivity of over 0.85 in 39 of the 41 labs that had at least 1 implausible observation. In the remaining two labs, Troponin I.cardiac (LOINC: 10839–9) and Cholesterol in LDL (LOINC: 13457–7), the best sensitivity was 0.0545 and 0.4867, respectively. Figure 4 illustrates the implausible values detected by the clustering method for the two labs. Troponin I.cardiac (LOINC: 10839–9) was an unusual case with a much higher proportion of implausible observations than expected (over 39,000 based on our silver standard implausible cutoff of [0–20] and normal range of [0.04–0.39]). Such a large number of implausible values violates our sparsity assumption under nonsystematic errors – i.e., the issue should be systematic and hence is observed frequently. For Cholesterol in LDL (LOINC: 13457–7) there were over 500 positives, or implausible observations, out of over 14 million lab records.
Comparing the clustering approach with conventional anomaly detection
The comparison of our proposed clustering approach to conventional anomaly detection (CAD) methods can be seen in Fig. 5. Overall, Fig. 5 shows that the clustering approach produced overwhelmingly better specificity than conventional anomaly detection. Best specificity from CAD methods was obtained from using 6 standard deviations as threshold for identifying outliers, which was outperformed by the clustering approach. It is important to notice that small differences among clustering results are meaningful as we deal with very large datasets for different observation types. For example, for a lab test with 10 million observations, a 0.001 difference in specificity between the clustering and CAD approaches means 10,000 more flagged observations that may need to be reviewed for plausibility.
In 31 of the 41 labs, the clustering and CAD produced similar sensitivity (Fig. 6). Among CAD methods, the best sensitivity was obtained from applying Mahalanobis Distances and 3.717526 (sqrt of 13.82) as critical value. The conventional anomaly detection (CAD) produced better sensitivity in 9 of the 41 labs for which we had implausible observations. The two largest delta in sensitivity was for Troponin T.cardiac (LOINC: 6598–7), where the clustering approach outperformed the best CAD result by 0.8794, and for Cholesterol in LDL (LOINC: 13457–7), where the best CAD result improved sensitivity by 0.5133. Outside these two labs, the average improvement in sensitivity in 9 labs was 0.1228, which considering the sparsity of implausible observations is minor.
We further evaluated the differences between CAD methods and our proposed clustering approach through number of false positive cases in each of the labs. As discussed earlier, our goal was to minimize the frequency of observations falsely identified as implausible. Figure 7 illustrates a pairwise comparison of number of false positive cases identified by each approach for each of the 50 lab observations. In 45 out of 50 labs (90%), the clustering approach produced a statistically significant smaller number of false positive cases.
More importantly, when the clustering approach outperformed the CAD approach, the gaps between the two approaches signified a large number of false positives. As the Yaxes show (scale is transformed in thousands) the conventional anomaly detection often identifies thousands of more plausible observations as implausible, compared with the clustering approach.
Discussion
EHRs provide massive amounts of observational data. Biological data have certain properties that are distinct in their distribution from other types of the so called “Big Data.” We designed, implemented, and tested an unsupervised clustering approach for identifying implausible records in clinical observation data. Our approach is based on two linked hypotheses that 1) if no systematic data entry errors exist, implausible clinical observations in electronic health records are sparse, and therefore 2) if clustered appropriately, clusters with very small populations should represent implausible observations. Using EHR laboratory results data from Partners Healthcare, our results supported both hypotheses. A set of plots on selected labs are available in Additional file 1: Figure S1.
We also demonstrated that the clustering approach outperforms conventional anomaly detection (CAD) approaches in identifying implausible lab observations. In biological data, an outlier may or may not be implausible. For example, if an EHR has a patient record for a 121yearold woman, the record is very likely an outlier, but not biologically implausible – according to the list of the verified oldest people in 1997, a French woman died at the purported age of 122 years, 164 days. Our results showed that, through the sparsity assumption (α) the clustering approach can improve differentiating the implausibles from the outliers, in comparison with the conventional methods. In anomaly detection, outliers are often conceived as extreme values at the two tails of the distribution. Our approach expands the conventional definition of an outlier, by searching for observations that are sparse considering their value and the density of the cluster they belong to, relative to the rest of the data points. Such observations can be found anywhere across the distribution of data – i.e., implausible values can still be extremely high or low, or just different enough from the rest of data points. For example, on a postpartum unit where ages may be ~ 0–4 days old or ~ 14–45 years old, one would not expect anyone 6 years old. In the 50 observation types we evaluated in this study, we did not find implausible values outsides of the upper or lower tails of the data distribution. However, labs and vital signs can be evaluated as multidimensional data by including various dimensions such as patients’ demographics and comorbidities, in which case a proper implementation of the clustering method can also detect and flag other types of implausible records.
The clustering approach for identifying implausible observations offers a precise “Big Data” solution for clinical and biological observations stored in electronic medical records. The clustering process is observationspecific, as the number of clusters and partitioning is specified for each group of observations. As a result, it produced low false positives – plausible observations mistakenly identified as implausible. In contrast, we showed that CAD approaches produce a high number of false positives. This difference is a huge benefit for the clustering approach from an informatics standpoint regarding implementation in large scale data repositories.
Limitations and directions for future research
We used a hybrid hierarchical Kmeans algorithm, HKmeans, because Kmeans algorithms are generally sensitive to outliers and the hybrid method improves Kmeans’ reliance on the random centroid initialization. We can imagine that other distancebased or densitybased unsupervised clustering algorithms might be also effective in identifying rare implausible clinical observations.
Many of clinical observations stored in EHR data are unlabeled. Unsupervised learning approaches offer many promising solutions for patient characterization. Nevertheless, these approaches require specification of several hyperparameter. In our case, performance of the clustering approach depended on the number of clusters, initial random assignment of cluster centroids, and the threshold for flagging clusters as implausible. In addition, implementation of the approach on very large datasets was challenging. To address some of these challenges, we had to be creative in selecting the clustering algorithm, applying feature transformations, and developing the kluster procedure [28] to approximate the number of clusters, all of which would be practical for future unsupervised learning efforts that aim to ascertain meaningful patient subgroups.
We randomly shuffled the data to prevent any potential sorting in breaking the data to folds for parallel computing. Further research is needed to ascertain whether some systematic approach to parallelization (e.g., breaking the data by age and gender) can improve the unsupervised implausible observation detection results.
Moreover, we have tested the clustering approach against 50 EHR observations. This demonstration was limited to a small set of observations due to the need for silverstandards to measure sensitivity and specificity of our proposed approach. However, we encourage the readers to envision further applications of this approach to other clinical observations, as well as complex combinations of observations in multiple dimensions, for which specification of manual silverstandards are virtually implausible.
Finally, the primary use case for this method is intended for an algorithmic screening of EHR laboratory observations for potential implausible values that would not be suggested for secondary use. The methodology can be applied to other use cases, such as identifying exceptionally wellperforming clinicians or detecting unusual patient management actions, would require identification of outlying but still physiologically plausible values. However, because the evaluation criteria may be different for other use cases (we focused on false positives for the detection of implausible observations in large scale EHR repositories), further work may be needed to adjust the α in order to optimize performance. For instance, in cases where outliers are more numerous and/or are closer to the remainder of the data, chart review would allow a more definitive (although not always completely definitive) determination of whether a value is clearly erroneous or not.
Implementation considerations
The algorithmic solution we presented in this paper proved as a feasible alternative for replacing the current manual rulebased procedures for identifying biologically implausible values. Given the size of data and the emphasis on sensitivity versus specificity, the choice of α (the ratio for flagging a cluster as implausible) can vary. In large clinical data repositories, an α between 1/4000 and 1/6000 would provide good balance between true positives and false negatives. A smaller α is also computationally more expensive. When the size of the dataset is small, a small α will be more appropriate. Nevertheless, the value for α can be adjusted over time to address the institutional needs and comfort level. Due to superb specificity, implementing the clustering approach offers a lowrisk solution to an expensive manual procedure that is hard to implement. We envision the proposed solution to constantly operate on the data base servers where EHR data are stored. We recommend that our algorithmic solution should be used to augment (rather than replace) human decisionmaking for improving quality of EHR data. After the implausible values are detected and flagged, a workflow is needed to initiate further actions needed in order to determine the density of the flagged observations. As we showed, because the flagged records are also sparse, the frequency of such flags should not be of concern.
For a specific lab test (Troponin I.cardiac – LOINC: 10839–9) we found an unusual high number of implausible observations (over 39,000), which was calculated based on our silver standard implausible cutoff of [0–20] and normal range of [0.04–0.39]. Further work is needed to fully discern the root cause for this issue in the data. Nearly all of these implausible values originated from a single type of Cardiac Troponin I test run at Brigham and Women’s hospital between 2001 and 2008. Although RPDR tries to normalize the units among labs assigned to the same LOINC code, it is possible that this particular cTI test is being reported with different units than the other tests in this LOINC code (for example, 0.41 could be reported routinely as 41). We are presently in discussion with the RPDR team about this possibility. Nonetheless, even in this case, all of the clustering algorithms produced results with 100% specificity, meaning that even when the sparsity assumption is not fulfilled, false positive cases will not be introduced.
Conclusion
Detecting implausible clinical observations in Electronic Health Record (EHR) data is a challenge, requiring availability of standards thresholds and rulebased procedures to query observations that are out of the plausibility range. Establishing rulebased procedures to address this task entails extensive hardcoding that would accumulate over time and dimensionality. We proposed an alternative viable algorithmic solution, using unsupervised clustering approach. The clustering approach is superior than conventional anomaly detection approaches and adaptable to different types of numerical EHR observation data.
Availability of data and materials
Since access to this dataset requires consent from IRB, the dataset cannot be made available publicly or upon request.
References
 1.
Brown JS, Kahn M, Toh S. Data quality assessment for comparative effectiveness research in distributed data networks. Med Care. 2013;51:S22–9. Available from: http://www.ncbi.nlm.nih.gov/pubmed/23793049.
 2.
Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. J Am Med Inform Assoc [Internet]. 2013;20:144–51 Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3555312/.
 3.
Kahn MG, Callahan TJ, Barnard J, Bauck AE, Brown J, Davidson BN, et al. A harmonized data quality assessment terminology and framework for the secondary use of electronic health record data. EGEMS (Wash DC). 2016;4(1):1244. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5051581/. [cited 2016 Sep 16].
 4.
Ghahramani Z. Unsupervised Learning. In: Bousquet O, von Luxburg U, Rätsch G, editors. Advanced Lectures on Machine Learning. ML 2003. Lecture Notes in Computer Science, vol 3176. Berlin, Heidelberg: Springer; 2004.
 5.
Hauskrecht M, Batal I, Hong C, Nguyen Q, Cooper GF, Visweswaran S, et al. Outlierbased detection of unusual patientmanagement actions: an ICU study. J Biomed Inform. 2016;64:211–21.
 6.
Bouarfa L, Dankelman J. Workflow mining and outlier detection from clinical activity logs. J Biomed Inform. 2012;45(6):1185–90.
 7.
Presbitero A, Quax R, Krzhizhanovskaya V, Sloot P. Anomaly detection in clinical data of patients undergoing heart surgery. Procedia Comput Sci. 2017;108:99–108.
 8.
Antonelli D, Bruno G, Chiusano S. Anomaly detection in medical treatment to discover unusual patient management. IIE Trans Healthc Syst Eng. 2013;3(2):69–77.
 9.
Ray S, Wright A. Detecting anomalies in alert firing within clinical decision support systems using anomaly/outlier detection techniques. Proc. 7th ACM Int. Conf. Bioinformatics, Comput. Biol. Heal. Informatics. New York: ACM; 2016. p. 185–90. Available from: http://doi.acm.org/10.1145/2975167.2975186
 10.
Ray S, McEvoy DS, Aaron S, Hickman TT, Wright A. Using statistical anomaly detection models to find clinical decision support malfunctions. J Am Med Informatics Assoc. 2018;25(7):862–71.
 11.
Wilson B, Tseng CL, Soroka O, Pogach LM, Aron DC. Identification of outliers and positive deviants for healthcare improvement: looking for high performers in hypoglycemia safety in patients with diabetes. BMC Health Serv Res. 2017;17(1):738.
 12.
Deneshkumar V, Senthamaraikannan K, Manikandan M. Identification of outliers in medical diagnostic system using data mining techniques. Int J Stat Appl. 2014;4(6):241–8.
 13.
Chandola V, Banerjee A, Kumar V. Anomaly detection: a survey. ACM Comput Surv 2009;41:1–58. Available from: https://dl.acm.org/citation.cfm?id=1541882.
 14.
Hodge VJ, Austin J. A survey of outlier detection methodologies. Artif Intell Rev. 2004;22(2):85–126.
 15.
Aggarwal CC, Yu PS. Outlier detection for high dimensional data. ACM SIGMOD Rec. 2001;30(2):37–46.
 16.
Knorr EM, Ng RT, Tucakov V. Distancebased outliers: algorithms and applications. VLDB J. 2000;8(34):237–53.
 17.
BenGal I. Outlier Detection. In: Maimon O, Rokach L, editors. Data Mining and Knowledge Discovery Handbook. Boston: Springer; 2005.
 18.
Gaspar J, Catumbela E, Marques B, Freitas A. A systematic review of outliers detection techniques in medical data  preliminary study. Heal. 2011. Proc Int Conf Heal Informatics. 2011.
 19.
Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: data mining, inference, and prediction: Springer Ser. Stat; 2009.
 20.
Jain AK. Data clustering: 50 years beyond Kmeans. Pattern Recogn Lett. 2010;31:651–66.
 21.
MacQueen J. Some methods for classification and analysis of multivariate observations. Proc. Fifth Berkeley Symp. Math. Stat. Probab. Vol. 1 Stat. Berkeley, Calif.: University of California Press; 1967. p. 281–97. Available from: http://projecteuclid.org/euclid.bsmsp/1200512992.
 22.
Chawla S, Gionis A. k means–: a unified approach to clustering and outlier detection. Proc. 2013 SIAM Int. Conf. Data min; 2013. p. 189–97. Available from: http://epubs.siam.org/doi/abs/10.1137/1.9781611972832.21.
 23.
Chen B, Tai PC, Harrison R, Pan Y. Novel hybrid hierarchicalKmeans clustering method (HKmeans) for microarray analysis. IEEE Comput Syst Bioinforma Conf Work Poster Abstr. 2005;2005:105–8.
 24.
Sugar CA, James GM. Finding the number of clusters in a dataset. J. Am. Stat. Assoc. 2003;98:750–63. Available from: http://www.tandfonline.com/doi/abs/10.1198/016214503000000666.
 25.
Hamerly G, Elkan C. Learning the k in k means. Adv neural Inf Process. 2004;17:1–8. Available from: books.nips.cc/papers/files/nips16/NIPS2003_AA36.pdf%5Cnhttp://books.google.com/books?hl=en&lr=&id=0F9C7K8fQ8C&oi=fnd&pg=PA281&dq=Learning+the+k+in+kmeans&ots=TGLvqYQa40&sig=SDu4cZ9TCeU8a5MoG1uMcRLQGFE.
 26.
Fraley C, Raftery AE. Modelbased clustering, discriminant analysis, and density estimation. J Am Stat Assoc. 2002;97:611–31. Available from: http://www.tandfonline.com/doi/abs/10.1198/016214502760047131.
 27.
Nalichowski R, Keogh D, Chueh HC, Murphy SN. Calculating the benefits of a research patient data repository. AMIA Annu Symp Proc United States. 2006. p. 1044.
 28.
Estiri H, Omran BA, Murphy SN. Kluster : an efficient scalable procedure for approximating the number of clusters in unsupervised learning. Big Data Res. 2018;13:38–51 Available from: http://linkinghub.elsevier.com/retrieve/pii/S2214579617303611. [cited 2018 Jun 3].
 29.
De Maesschalck R, JouanRimbaud D, Massart DLL. The Mahalanobis distance. Chemom Intell Lab Syst. 2000;50:1–18.
 30.
Filzmoser P. A multivariate outlier detection method. Seventh Int Conf Comput Data Anal Model. 2004.
Acknowledgements
The authors would like to thank Dr. Douglas S. Bell for validating the silverstandard plausibility thresholds.
Funding
This work was partially funded through a PatientCentered Outcomes Research Institute (PCORI) Award (CDRN130604608) for development of the National PatientCentered Clinical Research Network, known as PCORnet, NIH R01HG009174, and NLM training grant T15LM007092. The funding bodies had no role in the design of the study, collection, analysis, and interpretation of data or in writing the manuscript.
Author information
Affiliations
Contributions
HE conceptualized and designed the study, analyzed the data, interpreted the results, and drafted the paper. JK extracted the data for analysis and edited the paper. HE, JK, and SNM read and approved the final version of the paper.
Corresponding author
Correspondence to Hossein Estiri.
Ethics declarations
Ethics approval and consent to participate
This project was approved by the Partners Healthcare IRB with a waiver of consent according to protocol 2014P002367.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Additional file
Additional file 1:
Figure S1. Data distribution and implausible value detection for a set of selected EHR observation types.* xaxes are transformed to square root for visualization purpose. (PDF 369 kb)
Appendix
Appendix
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Received
Accepted
Published
DOI
Keywords
 Unsupervised clustering
 Implausible observations
 Data quality
 Electronic health records
 Informatics applications
 Anomaly detection