Professional Documents
Culture Documents
1 s2.0 S1532046421001696 Main
1 s2.0 S1532046421001696 Main
Original Research
A R T I C L E I N F O A B S T R A C T
Keywords: Electronic health records contain patient’s information that can be used for health analytics tasks such as disease
Electronic Health Records detection, disease progression prediction, patient profiling, etc. Traditional machine learning or deep learning
Image Mapping methods treat EHR entities as individual features, and no relationships between them are taken into consider
Convolutional Neural Networks
ation. We propose to evaluate the relationships between EHR features and map them into Procedures, Pre
Treatment Initiation Prediction
scriptions, and Diagnoses (PPD) tensor data, which can be formatted as images. The mapped images are then fed
into deep convolutional networks for local pattern and feature learning. We add this relationship-learning part as
a boosting module on a commonly used classical machine learning model. Experiments were performed on a
Chronic Lymphocytic Leukemia dataset for treatment initiation prediction. Experimental results show that the
proposed approach has better real world modeling performance than the baseline models in terms of prediction
precision.
1. Introduction predictions.
Recent works show incorporating hierarchical and sequential feature
Electronic health records (EHRs) collected through medical office relationships can help with improving the prediction performance. Choi
visits and pharmacy transactions can provide a significant amount of et al. proposed graph convolutional transformers that learn the hidden
digital information that can be used in applications such as health an structure of EHR while performing supervised training [10]. Choi et al.
alytics and clinical informatics. [1,2]. Various machine learning and have also proposed MiME [11], a model architecture that reflects re
deep learning models have been applied to EHR data for disease lationships between diagnosis and treatments. Wu et al. studied the
detection [3–6], disease or treatment progression [7,8], disease sub relationship between patients and clinicians because patients with the
typing classification [9], and so on. same disease that visit the same clinician tend to receive similar treat
In addition to information such as patient demographics: age, ments [12]. In EHR records, since feature interactions, mutual or con
gender, health care providers, and so on, an EHR also contains infor current effects can not be reflected automatically, researchers try to
mation relates to patient medical journeys: a sequence of diagnosis reconstruct those potential meaningful information using various
codes, surgical procedures, lab procedures, and prescriptions over time. methods. For example, the relationship between a prescription and a
Traditionally, the information in an EHR is treated as a flattened bag disease can be represented by how often they appear in the same med
of aggregated features or medical feature sequence in disease progres ical claim. Zeng et al. measured drug-drug similarity by calculating the
sion prediction models. In addition, each patient’s visit consists of Jaccard similarity coefficient on EHR data [13].
different types of features. For example, diagnosis and lab results will Deep convolutional neural networks (CNNs) are a popular class of
affect healthcare provider’s decisions on prescriptions and treatment deep learning models, and they are mostly used to analyze visual im
procedures. Also, the frequency and sequence of treatments in the early agery. Images contain spatially coherent pixels. Neighboring pixels
stage of disease will impact the choices of treatment plans in the late share similar information. Convolution filters in a CNN model do
stage. Unfortunately, such latent interactions between multiple EHR convolution operations and extract features among neighboring pixels.
features are hard to leverage to make significant contributions for model Sharma et al. [14] have transformed non-image data such as gene
* Corresponding author.
E-mail address: guanhao.wei@iqvia.com (G. Wei).
https://doi.org/10.1016/j.jbi.2021.103840
Received 4 February 2021; Received in revised form 8 June 2021; Accepted 11 June 2021
Available online 15 June 2021
1532-0464/© 2021 Elsevier Inc. This article is made available under the Elsevier license (http://www.elsevier.com/open-access/userlicense/1.0/).
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
expressions and test data to images and used CNNs for extracting fea genomic data, etc. Some researchers map non-images data to images and
tures and subsequent prediction tasks. When mapping non-image data to use CNNs to produce higher classification performance. Xu et al. turned
images, the positions of the pixels are essential. If they are arbitrarily texts into binary codes by applying 1-d convolution to them [17]. Gao
arranged, it can harm the feature extraction and prediction et al. converted DNA sequences to binary code and used CNNs to predict
performances. polyadenylation sites [18]. Lyu et al. used CNNs for RNA-seq data. The
Inspired by previous works, we propose to map EHR features, mainly resulting images are constructed based on chromosome location [19].
Procedures, Prescriptions, and Diagnoses (PPD) records to images using Sharma et al. converted various datasets into images. Relationships
relational information among the features. Such EHR mapped image is between a set of features are analyzed, and similar features are placed
defined with the term PPD tensor for this particular use case. This way, close to each other. Crucial information can be learned through element
the feature relationships are taken into consideration as related features arrangements. On the other hand, if features are arbitrarily arranged,
are positioned as neighboring pixels inside the PPD tensors. A CNN the prediction performance can be badly influenced [14].
model can then be used to extract and learn from the mapped PPD
tensors. Specifically, we experiment on EHR data on Chronic Lympho 3. Method
cytic Leukemia (CLL) patients and make predictions on whether the next
line of treatments needs to be initiated. 3.1. Mapping electronic health records to PPD tensors
The contribution of our work is as follows:
We take the diagnosis and treatment (procedure and prescription)
• We use a convolutional neural network to learn the relationship features from EHR data and map them into PPD tensors for each patient.
among medical entities in EHR data and perform a supervised pre As a result, each patient will have a unique PPD tensor representing his/
diction task. her doctor visits journey. The resulting tensors have three ranks. The
• We propose a novel method that maps EHR patient journeys (diag diagnosis and treatment features will be positioned inside of a matrix (a
nosis, procedures, and prescriptions) to PPD tensors. The underlying rank-2 tensor), and the time information will make the 3rd rank of the
relations among the medical entities are taken into consideration. tensor. For example, if there are 144 diagnosis and treatment features in
• Deep learning models require balances among classes. In the actual total, they can be mapped as a 12 × 12 matrix without time information.
data, the number of positive samples is small. We design a novel EHR To map diagnosis-treatment features into PPD tensors, we need to
data augmentation method that upsamples the positive class. consider: the relationships between features, how to position them ac
• We utilize a hybrid machine learning model architecture that in cording to their relationships, how to normalize feature values to the
cludes a base machine learning model and a CNN boosting model. range of [0, 1]. We also have some time information in our dataset, and
The base machine model can learn additional features that are not we encoded that into the 3rd rank of PPD tensors.
mapped to PPD tensors. We first analyze the relationship between diagnosis and treatment
features by evaluating quantitatively how often two features appear
The rest of our work is organized as follows: Section 2 is related together in the same medical claim. Then, we treat each feature like a
work. Section 3 introduces our method: how EHR data can be mapped to pixel inside an image and optimize their locations inside the image
PPD tensors and then fed to Convolutional Neural Networks to learn matrix by considering their relationships. Related features are posi
local patterns. Section 4 describes experiment settings and results. tioned as adjacent pixels as much as possible. Since the range of a pixel
Finally, Section 5 includes the conclusion and possible future directions. value is [0, 1], 0 representing black, and 1 representing white, we also
normalize our features into this range. At last, we put time information,
2. Related work adding the 3rd rank to the PPD tensor. The following subsections discuss
in detail how diagnosis-treatment features are mapped into image
2.1. Deep learning models on EHR data tensors.
The ability to proactively identify potential changes in a patient’s 3.1.1. Relationships between procedure, prescription and diagnosis
treatment is a core research task of disease or treatment progression Electronic health records contain various types of information such
modeling. Various deep learning models have been developed for this as patient demographic information like age, gender, and ethnicity; and
purpose. Ma et al. [15] proposed a bidirectional RNN, which is informed patient encounter information such as diagnosis codes, procedures used,
on both past and future visits. Three attention mechanisms are used to and drugs prescribed. In addition, there exist interactive relationships
take into consideration the relationships among visits. Che et al. [7] between the procedure, prescription, and diagnosis features. For
designed an RNN architecture for personalized predictions of Parkin example, doctors usually give the procedure EKG based on the diagnosis
son’s disease. Dynamic temporal matching is used to learn the similarity of chest pain. Therefore, it means that chest pain and EKG are closely
between longitudinal patient records. Choi et al. [6] developed a two- related.
level predictive model to detect influential past visits and clinical vari In images, pixels that are physically close together share similar in
ables and improved model accuracy and interpretability. Choi et al. [3] formation. Similarly, transforming Electronic Health Records into PPD
also proposed an RNN model for early detection of heart failures. Hi tensors requires positioning similar entities together. Electronic health
erarchical information from medical oncology is extracted using a records often do not contain direct relationships and closeness between
graph-based attention model. Ma et al. [16] proposed a framework for key medical entities. However, such information can be measured using
diagnosis prediction based on patients’ historical visit records. Medical quantitative methods. To extract these kinds of associations from EHRs,
code descriptions are incorporated in their model. we look at how frequently two medical entities appear together in a
medical claim. The more frequently they appear together, the more
likely they are related. Specifically, we use conditional probability to
2.2. Mapping non-image data to images measure the relationship between a treatment (procedure or prescrip
tion) and a diagnosis. Formula 1 shows how the relationship is
CNNs are deep learning models that handle images effectively. Many measured:
data are not in the form of images, such as texts, speeches, financial data,
2
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
3
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
4
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
and flipping do not suit our needs here. Since pixels of different co
ordinates have different meanings – they represent different diagnoses
or treatments, we do not want to use translations or rotations because
they change pixel locations.
A popular method used for EHR data augmentation is Synthetic
Minority Oversampling Technique (SMOTE) [29]. SMOTE works by
selecting two samples, drawing a line between them in the feature space,
and generating a new one along the line. Unfortunately, we found the
data augmentation using SMOTE to be very slow due to a large amount
of data and high feature dimensions that we have. We use a novel data
augmentation technique named PPD-tensor-overlay for EHR data. The
technique is inspired by data augmentation methods using image
overlay [30] As shown in Fig. 7, we take all features from a positive
sample and overlay it with part/all features from a negative one to
produce a new synthesized positive sample. Note that before the data
augmentation step, we have removed some negative training samples
that could potentially be positive using the information obtained from
the base machine model. The negative samples used by the data
augmentation is a reduced set of the original after the removal. Intui
tively, positive samples plus features from the negative samples should
still generate positive ones. This technique is easy to implement and runs
very efficiently in practice.
Formula 2 is used to generate a synthesized positive patient PPD Fig. 5. A 3 × 3 neighborhood example.
tensor.
Iaug [i, j] = min(1, Ipos [i, j] + Mask[i, j]⋅Ineg [i, j]) (2)
EHR data for every patient is processed, and the diagnosis and
treatment (prescription and procedure) information is converted to PPD
tensors. Thus, every patient has one image that encodes the EHR in
formation over a time interval. After mapping diagnosis and treatments
to PPD tensors, we pass them to a CNN for feature extraction and clas
sification. A CNN model of multiple convolution layers is created for a
binary classification task. Algorithm 2 describes how the convolution
filters learn the similarity information from neighboring pixels. Fig. 6. Time information encoded in the 3rd rank of a PPD tensor.
Algorithm 2. Convolution for feature similarity learning
5
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
Fig. 7. To generate a synthesize positive patient PPD tensor, a positive patient’s feature is overlaid with part of a negative patient’s features.
4. Experiments and results regarding the product, provider, payer, and geography. Rx data is
longitudinally linked back to an anonymous patient token and can be
We perform our experiments on a Chronic Lymphocytic Leukemia linked to events within the data set itself and across other patient data
dataset. We compare our results with base machine learning models on assets. In addition, IQVIA receives just under 1 billion office-based
the validation dataset. Top k patients with the highest scores are electronic medical claims from office-based individual professionals,
selected, and precisions are calculated for different k values. ambulatory and general healthcare sites per year, including patient-
level diagnosis and procedure information. The information represents
4.1. Chronic Lymphocytic Leukemia dataset nearly 40% of all electronically filed medical claims in the US and is now
the largest diagnosis database of its kind. The Dx data includes infor
We perform our experiments on the Chronic Lymphocytic Leukemia mation tracked to a specific prescriber enabling the user to obtain a more
patient dataset to make predictions for the second-line treatment (L2+) thorough picture of a prescriber’s activity which can, in turn, be linked
initiation in a fixed time window. This dataset is obtained from the to prescription data, affiliations, and other practitioner attributes.
IQVIA longitudinal prescription claims database (Rx) and electronic Information for patients who were diagnosed and received initial
medical claims (Dx) database. IQVIA receives 2 billion prescription treatment with CLL L2 + were selected from the databases. Patients
claims per year with history from January 2004 with coverage up to labeled as positive are the ones who received second-line treatment in
88% for the retail channel, 50–70% for traditional and specialty mail the 15 months look-back period interval. In contrast, negative ones were
orders, and 40% for long-term care. This information represents activ not observed to receive any further treatments inside of the interval. The
ities during the prescription transaction and contains information training dataset consists of monthly aggregated cohorts from April 2018
through May 2019. In addition, data from 4 cohorts from September
2019 to December 2019 are used for model validation. Due to potential
patient information overlap from neighboring cohorts, we ensured that
validation cohorts are at least three months later than the most recent
training cohort to avoid information leakage. Overall, there will be
around 60 K patients in each cohort. Thus, the amount of positive
Fig. 8. The overall architecture of the proposed modeling pipeline. Fig. 9. Overall model performance in terms of precision at top k.
6
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
Table 1
Average model prediction precision comparison at top k scored CLL patients.
Top k value 100 300 500 1,000 2,000 3,000
7
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
In Table 1, we show the precision@k values for a selected range of k Declaration of Competing Interest
values: 100, 300, 500, 1000, 2000 and 3000. We compare our method
with the base machine learning models: Random Forest, XGBoost, and The authors declare that they have no known competing financial
CatBoost. From Table 1, it can be observed that our CNN boosted ML interests or personal relationships that could have appeared to influence
model has overall better precision@k compared with the corresponding the work reported in this paper.
single baseline model. When the k value is smaller, the improvements
tend to be larger. In addition to listing single model performance, we References
also run experiments with combinations of base ML models. We can see
that a simple combination of RF, XGB, and CatBoost does not necessarily [1] Eric J. Topol. High-performance medicine: the convergence of human and artificial
intelligence. Nature Medicine, 25(1):44–56, 1 2019.
improve model performances, which also demonstrates the boosting
[2] Cao Xiao, Edward Choi, Jimeng Sun, Opportunities and challenges in developing
ability from the proposed CNN method. deep learning models using electronic health records data: a systematic review,
What’s more, to better specify the impact from the genetic algorithm Journal of the American Medical Informatics Association: JAMIA 25 (10) (2018)
and data augmentation of the proposed boosting approach, the model 1419–1428.
[3] Edward Choi, Andy Schuetz, Walter F Stewart, Jimeng Sun, Using recurrent neural
performances without optimized PPD tensor location (random mapping) network models for early detection of heart failure onset, Journal of the American
and using the SMOTE data augmentation technique are attached. It Medical Informatics Association: JAMIA 24 (2) (2017) 361–370.
shows that the genetic algorithm for feature position optimization is [4] Kezi Yu, Yunlong Wang, Yong Cai, Cao Xiao, Emily Zhao, Lucas Glass, and Jimeng
Sun. Rare Disease Detection by Sequence Modeling with Generative Adversarial
critically important for generating a high-quality PPD tensor that can be Networks. arXiv preprint arXiv:1907.01022, 2019.
fed into CNN models. A random set of feature positions does not [5] Zachary C Lipton, David C Kale, Charles Elkan, Randall P Wetzel Laura, and Leland
contribute much to generate a meaningful image for learning. Data K Whittier Virtual PICU. Learning to diagnose with lstm recurrent neural networks.
arXiv preprint arXiv:1511.03677, 2015.
augmentation is a crucial component of our CNN model part. Our CNN [6] Edward Choi, Mohammad Taha Bahadori, Joshua A Kulas, Andy Schuetz, Walter
model failed to work without data augmentation and cannot learn F Stewart, Jimeng Sun, RETAIN: An Interpretable Predictive Model for Healthcare
anything from the minority class since the two classes are highly un using Reverse Time Attention Mechanism, in: In Advances in Neural Information
Processing Systems, 2016.
balanced. SMOTE can also be used for minority class upsampling. Our
[7] Chao Che, Cao Xiao, Jian Liang, Bo Jin, Jiayu Zho, and Fei Wang. An RNN
PPD-tensor-overlay data augmentation runs much faster than SMOTE. Architecture with Dynamic Temporal Matching for Personalized Predictions of
We also include the SMOTE upsampling technique to balance the classes Parkinson’s Disease, in: Proceedings of the 2017 SIAM International Conference on
Data Mining, Philadelphia, PA, 6 2017. Society for Industrial and Applied
for reference. Our PPD-tensor-overlay data augmentation technique
Mathematics.
gives slightly higher model performances compared with SMOTE. [8] Fan Zhang, Tong Wu, Yunlong Wang, Yong Cai, Cao Xiao, Emily Zhao, Lucas Glass,
Compared to our production method, which is a fine-tuned XGB and and Jimeng Sun. Predicting Treatment Initiation from Clinical Time Series Data via
RF model ensemble, the CNN boosted XGB model improves absolute Graph-Augmented Time-Sensitive Model. arXiv preprint arXiv:1907.01099, 2019.
[9] Inci M Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K Jain, and Jiayu Zhou. Patient
precision at 1000 with 2.5%. It means many more truly identified pa Subtyping via Time-Aware LSTM Networks, in: Proceedings of the 23rd ACM
tients could potentially receive targeted treatment on time, bringing SIGKDD international conference on knowledge discovery and data mining, 2017.
significant value to drive the effectiveness of the overall treatment [10] Edward Choi, Zhen Xu, Yujia Li, Michael Dusenberry, Gerardo Flores, Emily Xue,
and Andrew Dai. Learning the graphical structure of electronic health records with
strategy. graph convolutional transformer, in: Proceedings of the AAAI Conference on
Artificial Intelligence, volume 34, pages 606–613, 2020.
5. Conclusion and future work [11] Edward Choi, Cao Xiao, Walter F Stewart, and Jimeng Sun. MiME: Multilevel
Medical Embedding of Electronic Health Records for Predictive Healthcare. arXiv
preprint arXiv:1810.09593, 2018.
In this work, we propose to map procedure, prescription, and diag [12] Tong Wu, Yunlong Wang, Yue Wang, Emily Zhao, Yilian Yuan, and Zhi Yang.
nosis features in EHR records to PPD tensors and use a CNN algorithm to Representation Learning of EHR Data via Graph-Based Medical Entity Embedding.
arXiv preprint arXiv:1910.02574, 2019.
learn the image features for model prediction boosting. The interactive
[13] Xian Zeng, Zheng Jia, Zhiqiang He, Weihong Chen, Xudong Lu, Huilong Duan, and
relationship between these medical features is taken into consideration. Haomin Li. Measure clinical drug–drug similarity using Electronic Medical
We apply our method to real-world healthcare record data, and exper Records. International Journal of Medical Informatics, 124(May 2018):97–103,
2019.
imental results show that it can significantly boost classical base ma
[14] Alok Sharma, Edwin Vans, Daichi Shigemizu, Keith A. Boroevich,
chine learning models’ performance. In the future, we plan to apply the Tatsuhiko Tsunoda, DeepInsight: A methodology to transform a non-image data to
new model design to various therapeutic areas and different clinic an image for convolution neural network architecture, Scientific Reports 9 (1)
scopes utilizing a similar data structure. In addition to full-image-based (2019) 1–7.
[15] Fenglong Ma, Radha Chitta, Jing Zhou, Quanzeng You, Tong Sun, and Jing Gao.
classification, recent studies as Recurrent Attention Convolutional Dipole: Diagnosis prediction in healthcare via attention-based bidirectional
Neural Network and Multiple Object Recognition have become hot recurrent neural networks, in: Proceedings of the 23rd ACM SIGKDD international
topics related to analyzing local information of an image and empha conference on knowledge discovery and data mining, pages 1903–1911, 2017.
[16] Fenglong Ma, Yaqing Wang, Houping Xiao, Ye Yuan, Radha Chitta, Jing Zhou, and
sizing such subpattern’s important role in the task. With the support of Jing Gao. A General Framework for Diagnosis Prediction via Incorporating Medical
those cutting-edge findings, we can further investigate identifying a Code Descriptions, in: 2018 IEEE International Conference on Bioinformatics and
subpattern of PPD feature combinations that make the greatest contri Biomedicine (BIBM), pages 1070–1075. IEEE, 12 2018.
[17] Jiaming Xu, Peng Wang, Guanhua Tian, Bo Xu, Jun Zhao, Fangyuan Wang, and
butions to model prediction performance, which can be essential refer Hongwei Hao. Convolutional Neural Networks for Text Hashing, in: IJCAI 15
ences for healthcare providers. Additionally, those medical features can Proceedings of the 24th International Conference on Artificial Intelligence, pages
be mapped to 3D images where more detailed time sequence informa 1369–1375, 12 2015.
[18] Xin Gao, Jie Zhang, Zhi Wei, Hakon Hakonarson, Deeppolya: A convolutional
tion can be encoded. The 3D images can then be fed into 3D-CNN models
neural network approach for polyadenylation site prediction, IEEE Access 6 (2018)
for feature learning. 24340–24349.
[19] Boyu Lyu and Anamul Haque. Deep Learning Based Tumor Type Classification
Using Gene Expression Data, in: BCB ’18: Proceedings of the 2018 ACM
6. CRediT authorship contribution statement
International Conference on Bioinformatics, Computational Biology, and Health
Informatics, pages 89–96, 2018.
Xueli Xiao: Methodology, Software, Visualization, Writing - Original [20] Steven R. Young, Derek C. Rose, Thomas P. Karnowski, Seung-Hwan Lim, and
Draft. Guanhao Wei: Conceptualization, Validation,Writing - Review & Robert M. Patton. Optimizing deep learning hyper-parameters through an
evolutionary algorithm, in: Proceedings of the Workshop on Machine Learning in
Editing. Li Zhou: Supervision, Writing - Review & Editing. Yi Pan: High-Performance Computing Environments - MLHPC ’15, pages 1–5, Austin,
Methodology, Resources. Huan Jing: Data Curation. Emily Zhao and Texas, USA, 2015. ACM Press.
Yilian Yuan: Project administration, Funding acquisition. [21] Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon
Suematsu, Jie Tan, Quoc Le, and Alex Kurakin. Large-Scale Evolution of Image
Classifiers, in: Proceedings of the 34th International Conference on Machine
Learning - Volume 70, pages 2902–2911, Sydney, NSW, Australia, 2017.
8
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840
[22] Xueli Xiao, Ming Yan, Sunitha Basodi, Chunyan Ji, and Yi Pan. Efficient [28] Connor Shorten and Taghi M. Khoshgoftaar. A survey on Image Data Augmentation
Hyperparameter Optimization in Deep Learning Using a Variable Length Genetic for Deep Learning. Journal of Big Data, 6(1):60, 12 2019.
Algorithm. arXiv preprint arXiv:2006.12703, 2020. [29] Nitesh V. Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer.
[23] Jatinder N.D Gupta and Randall S Sexton. Comparing backpropagation with a SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial
genetic algorithm for neural network training. Omega, 27(6):679–684, 12 1999. Intelligence Research, 16:321–357, 2002.
[24] David J Montana and Lawrence Davis. Training Feedforward Neural Networks [30] Hiroshi Inoue. Data Augmentation by Pairing Samples for Images Classification.
Using Genetic Algorithms, in: Proceedings of the 11th international joint arXiv preprint arXiv:1801.02929, 2018.
conference on Artificial Intelligence, pages 762–767, San Francisco, CA, USA, [31] T.K. Bamunu Mudiyanselage, X. Xiao, Y. Zhang, Y. Pan, Deep Fuzzy Neural
1989. Morgan Kaufmann Publishers Inc. Networks for Biomarker Selection for Accurate Cancer Detection, IEEE Trans.
[25] Xiuchun Xiao, Bing Liu, Jing Zhang, Xueli Xiao, and Yi Pan. An Optimized Method Fuzzy Syst., 2019.
for Bayesian Connectivity Change Point Model. Journal of Computational Biology, [32] Junbo Zhang, Jian-Syuan Wong, Tianrui Li, and Yi Pan. A comparison of parallel
25(3):337–347, 3 2018. large-scale knowledge acquisition using rough set theory on different MapReduce
[26] Xiuchun Xiao, Bing Liu, Jing Zhang, Xueli Xiao, and Yi Pan. Detecting Change runtime systems. International Journal of Approximate Reasoning, 55(3):896–907,
Points in fMRI Data via Bayesian Inference and Genetic Algorithm Model, in: 3 2014.
ISBRA 2017: Bioinformatics Research and Applications, pages 314–324. Springer, [33] Tianqi Chen and Carlos Guestrin. XGBoost: A Scalable Tree Boosting System, in:
Cham, 2017. Proceedings of the 22nd acm sigkdd international conference on knowledge
[27] Barrie M. Baker and M.A. Ayechew. A Genetic Algorithm for the Vehicle Routing discovery and data mining, pages 785–794, 2016.
Problem. Computers & Operations Research, 30(5):787–800, 4 2003.