Journal of Biomedical Informatics 120 (2021) 103840

Contents lists available at ScienceDirect

Journal of Biomedical Informatics

journal homepage: www.elsevier.com/locate/yjbin

Original Research

Treatment initiation prediction by EHR mapped PPD tensor based

convolutional neural networks boosting algorithm
Xueli Xiao a, Guanhao Wei b, *, Li Zhou b, Yi Pan a, Huan Jing b, Emily Zhao b, Yilian Yuan b
a
Computer Science Department, Georgia State University, Atlanta, GA 30303, USA
b
Advance Analytics, IQVIA Inc., Plymouth Meeting, PA 19462, USA

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
Electronic health records (EHRs) collected through medical office
visits and pharmacy transactions can provide a significant amount of
digital information that can be used in applications such as health an­
alytics and clinical informatics. [1,2]. Various machine learning and
deep learning models have been applied to EHR data for disease
detection [3–6], disease or treatment progression [7,8], disease sub­
typing classification [9], and so on.
In addition to information such as patient demographics: age,
gender, health care providers, and so on, an EHR also contains infor­
mation relates to patient medical journeys: a sequence of diagnosis
codes, surgical procedures, lab procedures, and prescriptions over time.
Traditionally, the information in an EHR is treated as a flattened bag
of aggregated features or medical feature sequence in disease progres­
sion prediction models. In addition, each patient’s visit consists of
different types of features. For example, diagnosis and lab results will
affect healthcare provider’s decisions on prescriptions and treatment
procedures. Also, the frequency and sequence of treatments in the early
stage of disease will impact the choices of treatment plans in the late
stage. Unfortunately, such latent interactions between multiple EHR
features are hard to leverage to make significant contributions for model
predictions.
Recent works show incorporating hierarchical and sequential feature
relationships can help with improving the prediction performance. Choi
et al. proposed graph convolutional transformers that learn the hidden
structure of EHR while performing supervised training [10]. Choi et al.
have also proposed MiME [11], a model architecture that reflects re­
lationships between diagnosis and treatments. Wu et al. studied the
relationship between patients and clinicians because patients with the
same disease that visit the same clinician tend to receive similar treat­
ments [12]. In EHR records, since feature interactions, mutual or con­
current effects can not be reflected automatically, researchers try to
reconstruct those potential meaningful information using various
methods. For example, the relationship between a prescription and a
disease can be represented by how often they appear in the same med­
ical claim. Zeng et al. measured drug-drug similarity by calculating the
Jaccard similarity coefficient on EHR data [13].
Deep convolutional neural networks (CNNs) are a popular class of
deep learning models, and they are mostly used to analyze visual im­
agery. Images contain spatially coherent pixels. Neighboring pixels
share similar information. Convolution filters in a CNN model do
convolution operations and extract features among neighboring pixels.
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.
expressions and test data to images and used CNNs for extracting fea­
tures and subsequent prediction tasks. When mapping non-image data to
images, the positions of the pixels are essential. If they are arbitrarily
arranged, it can harm the feature extraction and prediction
performances.
Inspired by previous works, we propose to map EHR features, mainly
Procedures, Prescriptions, and Diagnoses (PPD) records to images using
relational information among the features. Such EHR mapped image is
defined with the term PPD tensor for this particular use case. This way,
the feature relationships are taken into consideration as related features
are positioned as neighboring pixels inside the PPD tensors. A CNN
model can then be used to extract and learn from the mapped PPD
tensors. Specifically, we experiment on EHR data on Chronic Lympho­
cytic Leukemia (CLL) patients and make predictions on whether the next
line of treatments needs to be initiated.
The contribution of our work is as follows:
• We use a convolutional neural network to learn the relationship
among medical entities in EHR data and perform a supervised pre­
diction task.
• We propose a novel method that maps EHR patient journeys (diag­
nosis, procedures, and prescriptions) to PPD tensors. The underlying
relations among the medical entities are taken into consideration.
• Deep learning models require balances among classes. In the actual
data, the number of positive samples is small. We design a novel EHR
data augmentation method that upsamples the positive class.
• We utilize a hybrid machine learning model architecture that in­
cludes a base machine learning model and a CNN boosting model.
The base machine model can learn additional features that are not
mapped to PPD tensors.
The rest of our work is organized as follows: Section 2 is related
work. Section 3 introduces our method: how EHR data can be mapped to
PPD tensors and then fed to Convolutional Neural Networks to learn
local patterns. Section 4 describes experiment settings and results.
Finally, Section 5 includes the conclusion and possible future directions.
2. Related work
2.1. Deep learning models on EHR data
The ability to proactively identify potential changes in a patient’s
treatment is a core research task of disease or treatment progression
modeling. Various deep learning models have been developed for this
purpose. Ma et al. [15] proposed a bidirectional RNN, which is informed
on both past and future visits. Three attention mechanisms are used to
take into consideration the relationships among visits. Che et al. [7]
designed an RNN architecture for personalized predictions of Parkin­
son’s disease. Dynamic temporal matching is used to learn the similarity
between longitudinal patient records. Choi et al. [6] developed a two-
level predictive model to detect influential past visits and clinical vari­
ables and improved model accuracy and interpretability. Choi et al. [3]
also proposed an RNN model for early detection of heart failures. Hi­
erarchical information from medical oncology is extracted using a
graph-based attention model. Ma et al. [16] proposed a framework for
diagnosis prediction based on patients’ historical visit records. Medical
code descriptions are incorporated in their model.
2.2. Mapping non-image data to images
CNNs are deep learning models that handle images effectively. Many
data are not in the form of images, such as texts, speeches, financial data,
2
Journal of Biomedical Informatics 120 (2021) 103840
genomic data, etc. Some researchers map non-images data to images and
use CNNs to produce higher classification performance. Xu et al. turned
texts into binary codes by applying 1-d convolution to them [17]. Gao
et al. converted DNA sequences to binary code and used CNNs to predict
polyadenylation sites [18]. Lyu et al. used CNNs for RNA-seq data. The
resulting images are constructed based on chromosome location [19].
Sharma et al. converted various datasets into images. Relationships
between a set of features are analyzed, and similar features are placed
close to each other. Crucial information can be learned through element
arrangements. On the other hand, if features are arbitrarily arranged,
the prediction performance can be badly influenced [14].
3. Method
3.1. Mapping electronic health records to PPD tensors
We take the diagnosis and treatment (procedure and prescription)
features from EHR data and map them into PPD tensors for each patient.
As a result, each patient will have a unique PPD tensor representing his/
her doctor visits journey. The resulting tensors have three ranks. The
diagnosis and treatment features will be positioned inside of a matrix (a
rank-2 tensor), and the time information will make the 3rd rank of the
tensor. For example, if there are 144 diagnosis and treatment features in
total, they can be mapped as a 12 × 12 matrix without time information.
To map diagnosis-treatment features into PPD tensors, we need to
consider: the relationships between features, how to position them ac­
cording to their relationships, how to normalize feature values to the
range of [0, 1]. We also have some time information in our dataset, and
we encoded that into the 3rd rank of PPD tensors.
We first analyze the relationship between diagnosis and treatment
features by evaluating quantitatively how often two features appear
together in the same medical claim. Then, we treat each feature like a
pixel inside an image and optimize their locations inside the image
matrix by considering their relationships. Related features are posi­
tioned as adjacent pixels as much as possible. Since the range of a pixel
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,
adding the 3rd rank to the PPD tensor. The following subsections discuss
in detail how diagnosis-treatment features are mapped into image
tensors.
3.1.1. Relationships between procedure, prescription and diagnosis
Electronic health records contain various types of information such
as patient demographic information like age, gender, and ethnicity; and
patient encounter information such as diagnosis codes, procedures used,
and drugs prescribed. In addition, there exist interactive relationships
between the procedure, prescription, and diagnosis features. For
example, doctors usually give the procedure EKG based on the diagnosis
of chest pain. Therefore, it means that chest pain and EKG are closely
related.
In images, pixels that are physically close together share similar in­
formation. Similarly, transforming Electronic Health Records into PPD
tensors requires positioning similar entities together. Electronic health
records often do not contain direct relationships and closeness between
key medical entities. However, such information can be measured using
quantitative methods. To extract these kinds of associations from EHRs,
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
measure the relationship between a treatment (procedure or prescrip­
tion) and a diagnosis. Formula 1 shows how the relationship is
measured:
X. Xiao et al.
Fig. 1. An example of how related features are positioned as neighboring pixels
inside of a PPD tensor.
P(rx|dx) = P(rx ∩ dx)/P(dx)
P represents probability. rx is a treatment, and dx is a diagnosis. We
first go through all medical claims in our dataset, obtain the probabili­
ties of every single diagnosis and the probability of a treatment and a
diagnosis appearing together. Then, we calculate the conditional prob­
ability of a treatment given a diagnosis. This reflects quantitative the
relationship between diagnosis and treatments.
Fig. 2. The genetic representation of feature positions inside of a matrix.
Fig. 3. Modified crossover operation.
Journal of Biomedical Informatics 120 (2021) 103840
3.1.2. Genetic Algorithms for optimizing entity positions
When mapping medical features into PPD tensors, we treat each
feature as a pixel, and we want the related features to be positioned as
neighboring pixels as much as possible. Fig. 1 shows an example of how
some medical features are positioned inside of an image. Related fea­
tures, such as Fever, IV fluid, and Advil, are placed as neighboring pixels.
There might not be enough features to fill the image, so some pixels are
”empty” and do not represent any features. The fixed feature locations
and feature closeness reflect the image structure.
How to arrange the positions of diagnosis and treatments in a PPD
tensor such that related entities are as close together as possible is an
optimization problem. Suppose the number of diagnosis and treatments
together is N, and diagnosis and treatments are mapped to an image of
size W⋅H, there will be (W⋅H)!/(W⋅H − N)! number of ways to arrange. To
traverse through each possible arrangement and evaluate how close
together we are positioning the medical features is a computationally
impossible task in practice. As the number of features increases, the
number of ways of positioning grows super-exponentially.
Genetic algorithms (GA) can be used to find solutions to optimization
problems efficiently. A genetic algorithm is a kind of evolutionary al­
gorithm inspired by the process of natural selection. It is commonly used
to generate high-quality solutions when the search space size is enor­
mous. GAs have been applied to various optimization problems such as
deep learning hyperparameter optimizations [20–22], neural network
weight optimizations [23,24], search problems in neuroscience [25,26],
vehicle routing problems [27], etc. Typically, in a GA, an initial popu­
(1) lation of solutions is randomly generated and then evolved to create
better solutions. Biologically inspired operators such as selection,
crossover, and mutation are applied to the solutions during the evolu­
tionary process. As a result, the algorithm allows better solutions to
survive into new generations and eliminate bad ones.
To use a genetic algorithm, we need to have genetic representations
for the solutions. For our problem, we unroll the image mapping to a
vector that contains diagnosis and treatments, as shown in Fig. 2. The
3
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840

evolutionary process until the fitness score of the best-discovered solu­

tion converges (our stopping condition).
Algorithm 1. Diagnosis Treatments Mapping Optimization

Fig. 4. Mutation operations swaps two values.

unrolled vectors are the genetic representation of image mapping solu­

tions. We create random solution vectors and apply genetic operators to
get better and better image mapping solutions.
We design modified genetic operators to generate new populations.
We still use selection, crossover, and mutation operations. However, the
original crossover and mutation operations can not be directly applied
for ordering diagnosis and treatment because each entity appears in the
chromosome for exactly one time. Using the original crossover and
mutation would cause the solutions to contain some entities multiple
times while missing other entities. Instead, we use modified crossover
and mutation operations. Fig. 3 shows our crossover operations. Two
chromosomes, Parent a and Parent b, are recombined to generate two
children. The children contain genes from both parents. To produce
Child a, the front part of the chromosome from Parent a (bounded by the
red box) is first taken. However, we cannot simply use the latter part of
Parent b. Instead, from Parent b, we remove the genes Parent a has
already passed on to the child, concatenate the remaining genes in order,
and give Child a (bounded by the orange boxes). Child b is produced
similarly to Child a. For mutation, two genes in the chromosome are
randomly chosen, and their positions are swapped. These modified ge­
3.1.3. Values of medical features
netic operations guarantee that the children inherit from both parents
The value for each pixel inside a PPD tensors ranges from 0 to 1: zero
while containing each diagnosis and treatment for exactly one time. (See represents black, one means white, and the values in between are
Fig. 4).
various degrees of grays. The value for the procedure, prescription, and
We also design a fitness function to evaluate how good are the so­ diagnosis feature is the number of times that feature happened within a
lutions. We have calculated the relationship between a diagnosis and a
fixed period for the patient. To normalize each feature to the range of [0,
treatment (procedure or prescription) according to 1. For each treat­ 1], we divide each feature by the value at 95% percentile. Any results
ment, we have obtained its relationship values with all diagnosis fea­
greater than one are shrunk to 1. We normalized using the 95% value
tures. To calculate the fitness of a feature arrangement solution, we first instead of the maximum value to handle outliers.
find a treatment’s N-nearest diagnosis neighbors {N1 , N2 , ..., Nn } ac­
cording to the relationship values, with N1 being the closest diagnosis 3.1.4. Time information as 3rd rank in a tensor
neighbor. The number of nearest neighbors to consider (n) is a hyper­
The EHR data contains patients’ visit information over some time.
parameter to set up. We look at 3-nearest neighbors for our dataset. In And we want to incorporate some time information in our PPD tensor.
the image mapping, we define a neighborhood for each treatment. The
So, for a given time interval, instead of treating it as one whole part, we
neighborhood is simply the adjacent pixels to the treatment pixel. Fig. 5 divide it into three parts, and how many times each feature appears
is an example of a neighborhood. The prescription Advil is bounded by
within each piece is counted. The counts are then normalized to a value
the red box and has size 3 × 3, which means we consider the surrounding in the [0, 1] range. The three parts of time intervals make the 3rd rank of
8 pixels to be this prescription’s neighbors. The neighborhood size is a
the PPD tensor and make the 3rd rank’s dimension 3. Note that the
hyperparameter that can take different values. The larger the neigh­ length of time interval used in our dataset is the same for all EHR data.
borhood, the less constraint we put in the feature ordering. Next, for
Fig. 6 shows the time dimension of our PPD tensor: the front matrix
each treatment i, we check if any diagnosis in {N1 , N2 , ..., Nn } is within corresponds to features earlier in time, the back matrix means later.
its neighborhood and assign a corresponding score Ti . The value of Ti is
determined by which diagnosis is within I’s neighborhood. For example,
if N1 is within i’s neighborhood, Ti takes on a small value such as 0.5. If 3.2. Data augmentation
N3 is within i’s neighborhood, Ti has a slightly larger value such as 0.8. If
no diagnosis is within I’s neighborhood, Ti is infinity. The fitness score of One problem for prediction tasks on EHRs is that the dataset is
a PPD tensor mapping is sum(Ti ) for all treatments i. We want to opti­ usually highly unbalanced. The number of positive samples is minimal
mize the mapping to have small a fitness score as possible. compared with the negatives ones. Having highly unbalanced data will
The overall steps of our modified genetic algorithm for diagnosis- cause deep learning models to perform poorly since it will be difficult for
treatment to PPD tensor mapping are as follows. First, we designed the model to learn from the minority class. Using data augmentation
our crossover and mutation operation to produce new solutions. [28] to boost the number of positives samples is a solution to data
Different strategies to adopt crossover and mutation to discover new imbalance.
mapping solutions can be explored in the future. We repeat the Popular image augmentation methods such as translation, rotation,

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)

Iaug is the augmented image, Ipos is a positive image, Ineg is a negative

image. i and j are coordinates for a pixel in the image. Mask[i, j] is a mask
for the pixel that takes either 0 or 1. This value is randomly generated
according to a preset probability. The function min takes the minimum
value between two values.

3.3. Convolutional Neural Networks for image mapping classification

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

3.4. Overall model architecture

Fig. 8 shows the overall architecture of the proposed modeling

pipeline. Our input data contains both patient demographic information
and patient journey information. This information is fed to the
commonly used classical machine learning models, and prediction
scores are produced. The patient journey information - sequences of
diagnosis, procedure, and prescription are mapped to PPD tensors using
optimized mapping rules. We then use our data augmentation method to
balance the positive and negative classes. The output scores from the
base machine learning model are also used to remove some negative
training samples that could potentially be confusing to the deep learning
model. Negative sample removal is an optional step. We adopted this
step for the dataset we experimented on because the negative class is
noisy and may contain potentially positive samples. Therefore we
removed some negative training samples with a higher probability of
being positive to avoid confusing the deep learning model. The PPD
tensors are then fed to a CNN model for prediction. The output scores
from the CNN model are finally combined with the base machine
learning model for a final prediction score.

5
X. Xiao et al.
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
We perform our experiments on a Chronic Lymphocytic Leukemia
dataset. We compare our results with base machine learning models on
the validation dataset. Top k patients with the highest scores are
selected, and precisions are calculated for different k values.
4.1. Chronic Lymphocytic Leukemia dataset
We perform our experiments on the Chronic Lymphocytic Leukemia
patient dataset to make predictions for the second-line treatment (L2+)
initiation in a fixed time window. This dataset is obtained from the
IQVIA longitudinal prescription claims database (Rx) and electronic
medical claims (Dx) database. IQVIA receives 2 billion prescription
claims per year with history from January 2004 with coverage up to
88% for the retail channel, 50–70% for traditional and specialty mail
orders, and 40% for long-term care. This information represents activ­
ities during the prescription transaction and contains information
Fig. 8. The overall architecture of the proposed modeling pipeline.
6
Journal of Biomedical Informatics 120 (2021) 103840
regarding the product, provider, payer, and geography. Rx data is
longitudinally linked back to an anonymous patient token and can be
linked to events within the data set itself and across other patient data
assets. In addition, IQVIA receives just under 1 billion office-based
electronic medical claims from office-based individual professionals,
ambulatory and general healthcare sites per year, including patient-
level diagnosis and procedure information. The information represents
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
mation tracked to a specific prescriber enabling the user to obtain a more
thorough picture of a prescriber’s activity which can, in turn, be linked
to prescription data, affiliations, and other practitioner attributes.
Information for patients who were diagnosed and received initial
treatment with CLL L2 + were selected from the databases. Patients
labeled as positive are the ones who received second-line treatment in
the 15 months look-back period interval. In contrast, negative ones were
not observed to receive any further treatments inside of the interval. The
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. 9. Overall model performance in terms of precision at top k.
X. Xiao et al. Journal of Biomedical Informatics 120 (2021) 103840

patients who received second-line treatment is around 3% among all

CLL patients. In general, we extract Dx/Rx claim level data from the
IQVIA database to build features in different categories: patient char­
acteristics, major HCP information, diagnosis/prescription/treatment
counts, period gaps between various visit events, etc. Feature selection
methods using fuzzy logic or rough sets [31,32] could be explored in the
future to generate better features. In total, there are 226 features for
baseline models. And only 144 visit count features for each patient to
construct a mapped 12 × 12 PPD tensor.

4.2. Treatment initiation prediction model performance

Since our dataset is highly unbalanced between positives and nega­

Fig. 10. Precision Recall curve comparisons.
Fig. 11. CNN model architecture.
tives, we use precision@k to evaluate model performances. First, scores
are calculated for all patients, and k patients with the highest scores are
selected. We then calculate among the patients chosen what the per­
centage of patients who are actually positive is.
We apply our method to three baseline models for comparison:
Random Forest, XGBoost (XGB) [33], and CatBoost. We evaluate the
performance of these models with and without CNN learned feature
similarities. .
In a real-world application, considering limited time and resources
that can be put on truly identified patients for further healthcare
arrangement, we only care about model performance for the top 3000
patients with the highest scores. Thus we compare model precision for k
equals 3000 and less.
As shown in Figs. 9 and 10, we compare average model performances
on testing cohorts between baseline machine models with proposed CNN
boosted models. The ensemble of baseline models such as xgb-rf is
included as a reference. It can be seen that our model outperforms the
baselines in general.
Fig. 11 shows the architecture of the CNN we used in our model.
There are in total four convolutional layers with one fully connected
layer at the end. The hyperparameter settings can be found from the
figure.

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

Random Forest 24.0% 19.0% 17.4% 14.6% 12.5% 11.3%

CNN-Boost-RF 23.0% 19.3% 17.4% 14.9% 12.5% 11.3%

(random mapping, SMOTE) (-1.0%) (+0.3%) (+0.0%) (+0.3%) (+0.0%) (+0.0%)

CNN-Boost-RF 26.0% 20.6% 18.0% 15.8% 12.6% 11.3%

(optimized mapping, SMOTE) (+2.0%) (+1.6%) (+0.6%) (+1.2%) (+0.1%) (+0.0%)

CNN-Boost-RF 26.0% 21.0% 18.2% 15.8% 12.7% 11.3%

(optimized mapping, PPD-tensor-overlay) (+2.0%) (+2.0%) (+0.8%) (+1.2%) (+0.2%) (+0.0%)

XGB 27.0% 20.3% 17.8% 14.7% 12.7% 11.5%

CNN-Boost-XGB 27.0% 20.6% 18.2% 15.1% 12.7% 11.5%

(random mapping, SMOTE) (+0.0%) (+0.3%) (+0.4%) (+0.4%) (+0.0%) (+0.0%)

CNN-Boost-XGB 31.0% 23.0% 20.4% 17.0% 13.5% 11.6%

(optimized mapping, SMOTE) (+4.0%) (+2.7%) (+2.6%) (+2.3%) (+0.8%) (+0.1%)

CNN-Boost-XGB 33.0% 23.6% 21.0% 17.4% 13.7% 11.9%

(optimized mapping PPD-tensor-overlay) (+6.0%) (+3.3%) (+3.2%) (+2.7)% (+1.0%) (+0.4%)

CatBoost 25.0% 20.0% 17.8% 15.0% 12.7% 11.5%

CNN-Boost-CatB 24.0% 20.3% 18.0% 15.5% 12.9% 11.6%

(random mapping, SMOTE) (-1.0%) (+0.3%) (+0.2%) (+0.5%) (+0.2%) (+0.1%)

CNN-Boost-CatB 27.0% 22.3% 20.4% 16.9% 13.3% 12.0%

(optimized mapping, SMOTE) (+2.0%) (+2.3%) (+2.6%) (+1.9%) (+0.6%) (+0.5%)

CNN-Boost-CatB 27.0% 22.6% 20.4% 17.0% 13.5% 12.0%

(optimized mapping PPD-tensor-overlay) (+2.0%) (+2.6%) (+2.6%) (+2.0%) (+0.8%) (+0.5%)

RF + XGB 26.0% 19.6% 17.8% 14.9% 13.0% 11.7%

XGB + CatB 26.0% 20.3% 17.8% 14.8% 13.0% 11.8%

RF + CatB 25.0% 20.3% 17.4% 14.8% 12.8% 11.8%

7
X. Xiao et al.
In Table 1, we show the precision@k values for a selected range of k
values: 100, 300, 500, 1000, 2000 and 3000. We compare our method
with the base machine learning models: Random Forest, XGBoost, and
CatBoost. From Table 1, it can be observed that our CNN boosted ML
model has overall better precision@k compared with the corresponding
single baseline model. When the k value is smaller, the improvements
tend to be larger. In addition to listing single model performance, we
also run experiments with combinations of base ML models. We can see
that a simple combination of RF, XGB, and CatBoost does not necessarily
improve model performances, which also demonstrates the boosting
ability from the proposed CNN method.
What’s more, to better specify the impact from the genetic algorithm
and data augmentation of the proposed boosting approach, the model
performances without optimized PPD tensor location (random mapping)
and using the SMOTE data augmentation technique are attached. It
shows that the genetic algorithm for feature position optimization is
critically important for generating a high-quality PPD tensor that can be
fed into CNN models. A random set of feature positions does not
contribute much to generate a meaningful image for learning. Data
augmentation is a crucial component of our CNN model part. Our CNN
model failed to work without data augmentation and cannot learn
anything from the minority class since the two classes are highly un­
balanced. SMOTE can also be used for minority class upsampling. Our
PPD-tensor-overlay data augmentation runs much faster than SMOTE.
We also include the SMOTE upsampling technique to balance the classes
for reference. Our PPD-tensor-overlay data augmentation technique
gives slightly higher model performances compared with SMOTE.
Compared to our production method, which is a fine-tuned XGB and
RF model ensemble, the CNN boosted XGB model improves absolute
precision at 1000 with 2.5%. It means many more truly identified pa­
tients could potentially receive targeted treatment on time, bringing
significant value to drive the effectiveness of the overall treatment
strategy.
5. Conclusion and future work
In this work, we propose to map procedure, prescription, and diag­
nosis features in EHR records to PPD tensors and use a CNN algorithm to
learn the image features for model prediction boosting. The interactive
relationship between these medical features is taken into consideration.
We apply our method to real-world healthcare record data, and exper­
imental results show that it can significantly boost classical base ma­
chine learning models’ performance. In the future, we plan to apply the
new model design to various therapeutic areas and different clinic
scopes utilizing a similar data structure. In addition to full-image-based
classification, recent studies as Recurrent Attention Convolutional
Neural Network and Multiple Object Recognition have become hot
topics related to analyzing local information of an image and empha­
sizing such subpattern’s important role in the task. With the support of
those cutting-edge findings, we can further investigate identifying a
subpattern of PPD feature combinations that make the greatest contri­
butions to model prediction performance, which can be essential refer­
ences for healthcare providers. Additionally, those medical features can
be mapped to 3D images where more detailed time sequence informa­
tion can be encoded. The 3D images can then be fed into 3D-CNN models
for feature learning.
6. CRediT authorship contribution statement
Xueli Xiao: Methodology, Software, Visualization, Writing - Original
Draft. Guanhao Wei: Conceptualization, Validation,Writing - Review &
Editing. Li Zhou: Supervision, Writing - Review & Editing. Yi Pan:
Methodology, Resources. Huan Jing: Data Curation. Emily Zhao and
Yilian Yuan: Project administration, Funding acquisition.
8
Journal of Biomedical Informatics 120 (2021) 103840
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
References
[1] Eric J. Topol. High-performance medicine: the convergence of human and artificial
intelligence. Nature Medicine, 25(1):44–56, 1 2019.
[2] Cao Xiao, Edward Choi, Jimeng Sun, Opportunities and challenges in developing
deep learning models using electronic health records data: a systematic review,
Journal of the American Medical Informatics Association: JAMIA 25 (10) (2018)
1419–1428.
[3] Edward Choi, Andy Schuetz, Walter F Stewart, Jimeng Sun, Using recurrent neural
network models for early detection of heart failure onset, Journal of the American
Medical Informatics Association: JAMIA 24 (2) (2017) 361–370.
[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
Networks. arXiv preprint arXiv:1907.01022, 2019.
[5] Zachary C Lipton, David C Kale, Charles Elkan, Randall P Wetzel Laura, and Leland
K Whittier Virtual PICU. Learning to diagnose with lstm recurrent neural networks.
arXiv preprint arXiv:1511.03677, 2015.
[6] Edward Choi, Mohammad Taha Bahadori, Joshua A Kulas, Andy Schuetz, Walter
F Stewart, Jimeng Sun, RETAIN: An Interpretable Predictive Model for Healthcare
using Reverse Time Attention Mechanism, in: In Advances in Neural Information
Processing Systems, 2016.
[7] Chao Che, Cao Xiao, Jian Liang, Bo Jin, Jiayu Zho, and Fei Wang. An RNN
Architecture with Dynamic Temporal Matching for Personalized Predictions of
Parkinson’s Disease, in: Proceedings of the 2017 SIAM International Conference on
Data Mining, Philadelphia, PA, 6 2017. Society for Industrial and Applied
Mathematics.
[8] Fan Zhang, Tong Wu, Yunlong Wang, Yong Cai, Cao Xiao, Emily Zhao, Lucas Glass,
and Jimeng Sun. Predicting Treatment Initiation from Clinical Time Series Data via
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
Subtyping via Time-Aware LSTM Networks, in: Proceedings of the 23rd ACM
SIGKDD international conference on knowledge discovery and data mining, 2017.
[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
graph convolutional transformer, in: Proceedings of the AAAI Conference on
Artificial Intelligence, volume 34, pages 606–613, 2020.
[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.
[12] Tong Wu, Yunlong Wang, Yue Wang, Emily Zhao, Yilian Yuan, and Zhi Yang.
Representation Learning of EHR Data via Graph-Based Medical Entity Embedding.
arXiv preprint arXiv:1910.02574, 2019.
[13] Xian Zeng, Zheng Jia, Zhiqiang He, Weihong Chen, Xudong Lu, Huilong Duan, and
Haomin Li. Measure clinical drug–drug similarity using Electronic Medical
Records. International Journal of Medical Informatics, 124(May 2018):97–103,
2019.
[14] Alok Sharma, Edwin Vans, Daichi Shigemizu, Keith A. Boroevich,
Tatsuhiko Tsunoda, DeepInsight: A methodology to transform a non-image data to
an image for convolution neural network architecture, Scientific Reports 9 (1)
(2019) 1–7.
[15] Fenglong Ma, Radha Chitta, Jing Zhou, Quanzeng You, Tong Sun, and Jing Gao.
Dipole: Diagnosis prediction in healthcare via attention-based bidirectional
recurrent neural networks, in: Proceedings of the 23rd ACM SIGKDD international
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
Jing Gao. A General Framework for Diagnosis Prediction via Incorporating Medical
Code Descriptions, in: 2018 IEEE International Conference on Bioinformatics and
Biomedicine (BIBM), pages 1070–1075. IEEE, 12 2018.
[17] Jiaming Xu, Peng Wang, Guanhua Tian, Bo Xu, Jun Zhao, Fangyuan Wang, and
Hongwei Hao. Convolutional Neural Networks for Text Hashing, in: IJCAI 15
Proceedings of the 24th International Conference on Artificial Intelligence, pages
1369–1375, 12 2015.
[18] Xin Gao, Jie Zhang, Zhi Wei, Hakon Hakonarson, Deeppolya: A convolutional
neural network approach for polyadenylation site prediction, IEEE Access 6 (2018)
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
International Conference on Bioinformatics, Computational Biology, and Health
Informatics, pages 89–96, 2018.
[20] Steven R. Young, Derek C. Rose, Thomas P. Karnowski, Seung-Hwan Lim, and
Robert M. Patton. Optimizing deep learning hyper-parameters through an
evolutionary algorithm, in: Proceedings of the Workshop on Machine Learning in
High-Performance Computing Environments - MLHPC ’15, pages 1–5, Austin,
Texas, USA, 2015. ACM Press.
[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.
X. Xiao et al.
[22] Xueli Xiao, Ming Yan, Sunitha Basodi, Chunyan Ji, and Yi Pan. Efficient
Hyperparameter Optimization in Deep Learning Using a Variable Length Genetic
Algorithm. arXiv preprint arXiv:2006.12703, 2020.
[23] Jatinder N.D Gupta and Randall S Sexton. Comparing backpropagation with a
genetic algorithm for neural network training. Omega, 27(6):679–684, 12 1999.
[24] David J Montana and Lawrence Davis. Training Feedforward Neural Networks
Using Genetic Algorithms, in: Proceedings of the 11th international joint
conference on Artificial Intelligence, pages 762–767, San Francisco, CA, USA,
1989. Morgan Kaufmann Publishers Inc.
[25] Xiuchun Xiao, Bing Liu, Jing Zhang, Xueli Xiao, and Yi Pan. An Optimized Method
for Bayesian Connectivity Change Point Model. Journal of Computational Biology,
25(3):337–347, 3 2018.
[26] Xiuchun Xiao, Bing Liu, Jing Zhang, Xueli Xiao, and Yi Pan. Detecting Change
Points in fMRI Data via Bayesian Inference and Genetic Algorithm Model, in:
ISBRA 2017: Bioinformatics Research and Applications, pages 314–324. Springer,
Cham, 2017.
[27] Barrie M. Baker and M.A. Ayechew. A Genetic Algorithm for the Vehicle Routing
Problem. Computers & Operations Research, 30(5):787–800, 4 2003.
Journal of Biomedical Informatics 120 (2021) 103840
[28] Connor Shorten and Taghi M. Khoshgoftaar. A survey on Image Data Augmentation
for Deep Learning. Journal of Big Data, 6(1):60, 12 2019.
[29] Nitesh V. Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer.
SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial
Intelligence Research, 16:321–357, 2002.
[30] Hiroshi Inoue. Data Augmentation by Pairing Samples for Images Classification.
arXiv preprint arXiv:1801.02929, 2018.
[31] T.K. Bamunu Mudiyanselage, X. Xiao, Y. Zhang, Y. Pan, Deep Fuzzy Neural
Networks for Biomarker Selection for Accurate Cancer Detection, IEEE Trans.
Fuzzy Syst., 2019.
[32] Junbo Zhang, Jian-Syuan Wong, Tianrui Li, and Yi Pan. A comparison of parallel
large-scale knowledge acquisition using rough set theory on different MapReduce
runtime systems. International Journal of Approximate Reasoning, 55(3):896–907,
3 2014.
[33] Tianqi Chen and Carlos Guestrin. XGBoost: A Scalable Tree Boosting System, in:
Proceedings of the 22nd acm sigkdd international conference on knowledge
discovery and data mining, pages 785–794, 2016.