IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

Bayesian-Based Symptom Screening for Medical

Dialogue Diagnosis
Zhong Cao, Yuchun Guo∗ , Yishuai Chen, Daoqin Lin
School of Electronic and Information Engineering
Beijing Jiaotong University
Beijing, China
Email:{21111023,ychguo,yschen,20125050}@bjtu.edu.cn

Abstract—In a medical dialogue diagnosis system, the se- to predict a disease, do not get good classification results
lection of symptoms for inquiry has a significant impact when there is a high overlap of symptoms between diseases.
2023 IEEE Symposium on Computers and Communications (ISCC) | 979-8-3503-0048-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ISCC58397.2023.10218219

on diagnostic accuracy and dialogue efficiency. In a typical Recently there has been an increase in research based on
diagnosis process, the symptoms initially reported by users are
often insufficient to support an accurate diagnosis, making it deep reinforcement learning, [2] defined the query and diag-
necessary to ask users about other symptoms through dialogue nosis process as Markov decision process and using deep
to form a conclusive diagnosis. In this paper, we propose a Q-network (DQN) for policy learning. Based on this, [3]
disease diagnosis algorithm based on Bayesian, which simulates introduced knowledge graphs to add knowledge branch and
the process of doctor’s inquiry and diagnosis by dynamically knowledge-routed graph branch, but with limited enhance-
updating the list of diseases to increase the interpretability of
diagnosis results. For the symptom interrogation, we propose ment. However, RL-based methods have poor interpretability,
a symptom screening algorithm based on the difference of require more resources, and are also poorly transferable.
symptom sets to exclude diseases with low probability. Through
the intersection and union of disease symptom sets, we can In this paper, we propose Bayesian-based symptom screen-
screen out the symptoms that can distinguish diseases in fewer ing algorithm. We get the rough probabilities of diseases
inquiring rounds. The experimental results demonstrate the based on the positive and negative symptoms in the patient’s
proposed method performs more efficiently than existing state- self-report, using Bayesian and ranking them in descending
of-the-art algorithms. order. Symptom screening combines the variability between
Index Terms—Medical dialogue diagnosis system, symptom
screening, Bayesian inference, e-health disease symptom sets to identify the symptoms that best
distinguish diseases, we select high probability symptoms in
I. I NTRODUCTION low probability diseases to get more additional information.
Symptom screening is performed using a binary search,
The medical dialogue diagnosis system is one of the most which reduces the number of diseases by half in each round
important artificial intelligence applications in healthcare. of inquiry, greatly improving the efficiency of inquiry. The
The medical dialogue diagnosis system simulates a doctor, patient’s additional information and self-report are then used
inquiring the patient and combining their self-reports to to make the final disease prediction. Our contribution is as
predict potential disease. Medical dialogue diagnosis system follows:
can provide pre-consultation to patients and also help doctors
to collect information from patients. 1) We propose lightweight Bayesian-based symptom
The main challenges are that patients describe their symp- screening algorithm to predict patients’ diseases by multi-
toms incompletely and that one symptom can be associated ple rounds of interrogation. The main advantages are high
with multiple diseases, which makes diagnosis more difficult. interpretability, low computational complexity, and less data
The current research interest is to make a more accurate resources and computational resources required.
diagnosis by having multiple rounds of dialogue with the 2)We propose symptom screening algorithm. The variabil-
patient to make up for information that the patient is not ity between sets of disease symptoms is exploited to find the
actively giving. Also, using as few dialogue rounds as symptoms that best distinguish the disease. The number of
possible to check enough additional symptoms is desired for interrogation rounds is greatly reduced by binary search. The
the efficiency of this system. best results are obtained on two real-world datasets.
In traditional methods, [1] developed an expert system by
extracting knowledge from datasets and human knowledge, The rest of this paper is organized as follows. Section
but it has poor transferability. Machine learning approaches II reviews the related works. In Section III, we describes
such as decision tree, SVM, random forest have been used the overall framework and details of the algorithm. We then
presents the experimental setup and results in Section IV.
* Yuchun Guo is the corresponding author. Finally, we conclude in Section V.

979-8-3503-0048-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.
 IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

II. R ELATED WORK patient’s self-report. Bayesian inference is performed using

In the work of intelligent diagnosis, the first Bayesian these symptoms to give the probability of all diseases. The
inference and tree-based methods used entropy function symptom screening algorithm is used during the inquiry stage
to select symptoms based on information gain theory. [4] to complement the patient’s symptoms. We assume that the
and [5] proposed to address the cost of feature acquisition patient will answer the system’s questions based on his or
by using decision trees and random forest methods. [1] her real situation.
attempted to extract rule-based representations from medical
B. Notations
data and human knowledge for diagnosis. [6] and [7] sug-
gested that trained predictive factors for sensitivity analysis The disease set is represented by D, and the size of D is m.
to measure the importance of each feature in a given con- The symptom set is represented by S, S represents all symp-
text. [8] developed a hierarchical finding framework based toms of the disease d, where d ∈ D, S + and S − represent
on symptom embedding in which graphical representation the positive and negative symptom sets, respectively.
learning is customized for disease diagnosis. [9] proposed
C. Bayesian inference for disease diagnosis
approach based on the supervised learning of neural models
with logic regularization. These methods often use greedy Based on the distribution of the datasets, we can get the
or approximation algorithms due to the global maximization probability P (d = 1) of each disease being positive and the
of information gain or the recalcitrance of global sensitivity, probability P (s = 1|d = 1) of each symptom being positive
resulting in low accuracy. under each disease being positive,where s ∈ S .Given the
With the latest progress of deep reinforcement learning and symptom status information, the conditional probability of
task-oriented dialogue systems, the application research of each disease can be obtained as follows:
deep reinforcement learning [10] in medical dialogue systems P (d = 1, s = 1)
has gradually increased. In the context of symptom-based P (d = 1|s = 1) = (1)
P (s = 1)
dialogue system for disease diagnosis, [2] proposed a human-
machine dialogue system for automatic diagnosis. This au- where
tomatic diagnosis system defined the query and diagnosis X
P (s = 1) = P (s = 1|d = 1) ∗ P (d = 1) (2)
process as a Markov decision process, and applied a deep Q-
d∈D
Network (DQN) to learn the policy of the system. To decide
whether to continue the inquiry by asking about a symptom We cannot confirm that the disease d is negative when the
or to end the diagnosis by predicting a disease. It establishes symptom s is negative, so the probability that both are
a relatively complete dialogue system, which takes the first negative can be
step to solve the problem of automatic diagnosis of medical
P (d = 0|s = 0)
dialogue.
KR-DS [3] contains DQN based on knowledge path. By = 1 − P (d = 1|s = 0)
P (3)
considering the relationship between diseases and symptoms, P (s = 0|d = 1) ∗ P (d = 1)
= 1 − d∈D
the knowledge branch and knowledge routing branch are P (s = 0)
added to the deep reinforcement learning framework, and the
where
external probability symptoms related to the reinforcement
P (s = 0) = 1 − P (s = 1) (4)
learning framework are incorporated into it. It improves the
rationality of medical dialogue decision-making. according to equation 1-4 we can get
III. M ETHODOLOGY
In this section, we introduce the Bayesian-based symptom P (d = 1, S + , S − )
P (d = 1|S + , S − ) = (5)
screening algorithm. The probability distribution of symp- P (S + , S − )
toms and diseases in the dataset is used to derive the proba- where
bility of each disease of a patient by combining the explicit
P (d = 1, S + , S − )
and implicit symptoms. The symptom screening combines Y Y
the variability between disease symptom sets to identify the = P (d = 1) P (s = 1|d = 1) P (s = 0|d = 1)
symptoms that can best distinguish diseases. s∈S + s∈S −
Y Y
A. Task description = P (d = 1) P (s = 1|d = 1) (1 − P (s = 1|d = 1))
s∈S + s∈S −
In this work, we consider each patient has only one dis- (6)
ease,because it is quite common in some healthcare practice.
The symptom status is only negative or positive and the
X
P (S + , S − ) = P (d = 1, S + , S − ) (7)
system extracts negative and positive symptoms from the d∈D

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.
 IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

Note that the term P (d = 1, S + , S − ) can be factorized take situations with four diseases. Each circle represents the
because the symptom are conditionally independent given full set of symptoms of a disease, the slash area is the target
all diseases. symptom set, blue and orange represent the symptom sets
The Bayesian inference algorithm is summarized as Al- belonging to SH and SL , respectively.Our goal was to find
gorithm 1. Step 1 initializes an empty list DP to store the the symptom that best distinguished the disease. We put the
probabilities of all diseases. Step 2 calculates for each dis- symptom into S + or S − depending on the patient’s negative
ease, the joint probability of S + and S − . Step 3 calculates the or positive response at each inquiry, while removing half
probability of S + and S − . Step 4 calculates the conditional of the symptom set. Because the current results are biased
probabilities of each disease when the conditions of S + and towards high probability diseases, the system receives less
S − are all added to the DP . benefit if the symptoms are selected from the high probability
disease symptom set to ask the patient, in other words, the
Algorithm 1: Bayesian inference algorithm for diag- Bayesian inference process gets more additional information
nosis when the symptoms are selected from the low probability
Input: Positive symptom S + and negative symptom disease symptom set to ask the patient.
S−. In order to make the screened symptoms correlated with
Output: List DP of probability in each disease all sets in SL and uncorrelated with all sets in SH , we take
1 Initialize an empty list DP ; the union set of Si in SH :
2 for i in 1...m do [
3 Compute P (di = 1, S + , S − ) using equation 6; S∪ = Si (8)
Si ∈SH
4 end
+ − we take the intersect set of Si in SL :
5 Compute P (S , S ) using equation 7;
6 for i in 1...m do \
S∩ = Si (9)
7 Compute P (di = 1|S + , S − ) using equation 5;
Si ∈SL
8 add it to the list DP ;
9 end Then, we get the target set of symptoms:
\ [
SO = S∩ − S∪ = Si − Si (10)
Si ∈SL Si ∈SH
D. Symptom screening
Generally, patient do not provide enough information for After we get the target symptom set SO , in order to
the doctor to confirm the disease, so additional information is make the selected symptoms work in most cases, so we
required through multiple rounds of inquiry. We are inspired select the most frequently occurring symptoms according to
by this to propose symptom screening algorithm based on the probability distribution of the symptoms, which is the
the difference of symptom sets of each disease. following:
Patient self-reports were used to get rough disease proba- sOmax = argmax P (s = 1) (11)
s∈SO
bilities using Algorithm 1 and sorted in descending order(eq
P (d0 |S + , S − ) > P (d1 |S + , S − ) > ... > P (dm−1 |S + , S − ) So we select the symptom sOmax with the highest proba-
). Then we get the high probability disease symptom set bility of positive in the SO , add it to the positive symptom set
m S + or negative symptom set S − according to the patient’s
SH = {Si , i ∈ [1, 2...⌈ ⌉]} and the low probability disease
2 m answer.
symptom set SL = {Si , i ∈ [⌈ ⌉−1, ...m]}. Figure 1 shows Figure 2 represents the complete interrogation process
2
of the system when four diseases are present. The dashed
circles indicate the set of eliminated symptoms. In the first
S1 S3 round, the system gets the current sOmax symptoms by the
SH
symptom screening algorithm and interrogates the patient.
When sOmax is positive, S1 and S2 are eliminated and
SL sOmax is put into the positive symptom set S + ; when the
interrogation result of sOmax is negative, S3 and S4 are
SO eliminated and sOmax is put into the negative symptom set
S2 S4 S − . By the second round, take the left half as an example.
At this point, sOmax is recalculated, and when sOmax is
positive, S3 is eliminated and sOmax is put into the positive
symptom set S + ; the current when sOmax is negative, S4
Fig. 1. A round of symptom screening process
is eliminated and sOmax is put into the negative symptom
our symptom screening process for each round, selected to set S − . After the second round, only one disease symptom

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.
 IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

set remains, and this can be stopped. The final disease

𝑆1 𝑆3 probabilities are obtained by putting the updated S + and S −
𝑆H
using Algorithm 1.
sOmax is positive sOmax is negative 𝑆L The symptom screening is summarized as Algorithm 2.
𝑆2 𝑆4
𝑆O
Step 1 obtains the list of disease probabilities DP based
𝑆1 𝑆3
𝑆3 on Algorithm 1 and S + and S − in the patient’s self-report.
𝑆1 𝑆3
sOmax is positive sOmax is negative sOmax is positive sOmax is negative Step 2 calculates the length of the current disease list during
𝑆1 each interrogation, arranges the diseases of DP in descending
𝑆2 𝑆4 𝑆2 𝑆4 order of probability, and obtains sOmax after the symptom
screening algorithm. Step 3 interrogates the patient, adds
sOmax to S + if it is positive, and updates the disease
𝑆1 𝑆3 𝑆1 𝑆3 𝑆1 𝑆3 𝑆1 𝑆3
probability list to low-probability diseases. If negative, the
sOmax is added to S − and the disease probability list is
𝑆2 𝑆4 𝑆2 𝑆4 𝑆2 𝑆4 𝑆2 𝑆4 updated to high probability disease. When there is only one
disease left in the disease list, use Algorithm 1 to output the
disease with the highest probability.
Fig. 2. The symptom screening process in the four disease symptom sets
IV. E XPERIMENT

Algorithm 2: Symptom screening algorithm based on A. Dataset

Bayesian inference MZ [2] was collected on the China Baidu Thumb Doc-
Input: Positive symptom S + and negative symptom tor online consultation platform in 2018. It collected four
S − , maximum step N. common childhood diseases and contains 710 conversations
Output: pick the disease with the highest probability covering 66 symptoms of 4 diseases. The corresponding
in the list DP relationships between the four diseases and the number of
1 n = 0;
conversations and symptoms are shown in Table I.
2 Compute list DP of probability in each disease using DX [3] was collected from an online healthcare commu-
Algorithm 1; nity in China (Lilac Garden), which retained the original
3 while n ≦ N do self-report and interaction discourse. DX annotation for five
4 Calculate the length L of DP ; types of diseases, including allergic rhinitis, upper respiratory
5 if L ≦ 1 or n = N then tract infection, pneumonia, hand, foot and mouth disease in
6 Compute list DP of probability in each children and childhood diarrhea, [3] took in self-reported
disease using Algorithm 1; symptoms, and the conversation and normalized to 41 kinds
7 return of symptoms, shown in Table I.
8 else
9 n = n + 1; TABLE I
T HE N UMBER OF C ONVERSATIONS AND S YMPTOMS OF E ACH D ISEASE
10 Sort the disease probabilities in DP in
descending order; dataset disease dialogue symptoms
11 Get SH and SL according to DP ;
pediatric diarrhea 200 56
12 Get SO using equation 10; upper respiratory tract infection 160 59
MZ
13 Get sOmax using equation 11; minor bronchitis 200 61
14 if sOmax is positive then pediatric dyspepsia 150 66
15 Add sOmax to the positive symptom set allergic rhinitis 102 24
upper respiratory tract infections 122 23
S+ ; DX pneumonia 100 29
L
16 DP ← DP [⌈ ⌉ + 1, ..., L] ; Hand, foot and mouth disease 101 22
2 childhood diarrhea 102 33
17 else
18 Add sOmax to the negative symptom set
S− ; B. Baseline
L
19 DP ← DP [1, ..., ⌈ ⌉] ;
2 SVM treats disease diagnosis as a multi-class classification
20 end problem. It takes as input a one-hot vector representation
21 end of the patient’s symptoms and predicts disease. There are
22 end two configurations: one that takes both explicit and implicit
symptoms as inputs (denoted SVM-ex+im), and the other

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.
 IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

TABLE II
P ERFORMANCE COMPARISONS ON MZ DATASET.

method upper respiratory tract infection minor bronchitis pediatric diarrhea pediatric dyspepsia Overall Ave rounds
SVM-ex 0.44 0,71 0.89 0.28 0.59 -
SVM-ex+im 0.52 0.93 0.91 0.34 0.71 -
Basic DQN - - - - 0.65 -
KR-DS - - - - 0.73 -
Bayesian-ex 0.5333 0.8235 0.77 0.3733 0.6338 1
Bayesian-ex+im 0.5667 0.8235 0.8444 0.3939 0.6761 3.26
ours 0.833 0.9412 0.9778 0.909 0.9225 2.89

TABLE III
P ERFORMANCE COMPARISONS ON DX DATASET.

method allergic rhinitis upper respiratory pneumonia Hand, foot and mouth childhood diarrhea Overall Ave rounds
tract infections disease
Basic DQN - - - - - 0.731 3.92
KR-DS - - - - - 0.74 3.36
Bayesian-ex 0.6 0.625 0.55 0.9 1 0.7308 1
Bayesian-ex+im 0.95 0.3750 0.85 0.85 0.95 0.7788 2.67
ours 1 0.9167 0.85 1 1 0.9519 3.08

that uses only explicit symptoms to predict disease (denoted D. Result

SVM-ex). As shown in Tables II and III, the accuracy of Bayesian
Basic DQN defines the query and diagnosis process as a for most diseases in the MZ dataset is higher than SVM.
Markov decision process, and on this basis, deep Q-network When only explicit symptoms are used for disease diagnosis,
(DQN) is applied to learn the system policy-wise to decide Bayesian outperforms SVM in the classification of upper
whether to continue the conversation by asking about a respiratory tract infection, pediatric bronchitis, and pediatric
symptom or to end the diagnosis by predicting a disease. dyspepsia by approximately 9%. Additionally, the overall
KR-DS adds knowledge branches and knowledge routing accuracy of Bayesian increased by 4.38%. However, when
branches to the deep reinforcement learning framework by both explicit and implicit symptom information are used
considering the relationship between diseases and symptoms. for diagnosis, the accuracy of Bayesian was 3.39% lower
It improves the rationality of medical dialogue decision- than that of SVM, which showed a 22% improvement in
making. disease diagnosis accuracy. The experimental results proved
that Bayesian was better than that of SVM when only explicit
symptoms were used for diagnosis. However, Bayesian was
C. Setup similar to that of SVM when the disease was diagnosed
according to both explicit and implicit symptom information.
In terms of data set division, according to the division Both Basic DQN and KR-DS are based on the explicit
of data set by [2] and [3], training set: test set = 568:142 symptoms in the data set, and by learning the symptom rules
in MZ data set. In the DX dataset, training set: test set asked by doctors in the data set, they can screen symptoms
=423:104. Symptoms with ”UNK” symptom value were and make decisions. As shown in tableIII and II, compared
identified as invalid symptom information and the symptom with the best results, the disease diagnosis algorithm is
was eliminated.And only when the disease and symptom improved by 19.25% and 21.19%, respectively.
are both positive, the symptom can be added to the disease In the MZ dataset, the average dialogue rounds in the test
symptom set. The initial inquiry round number turn is 1, set is 4.20, and the average dialogue rounds of the disease
because after the user enters the user report, a full inquiry diagnosis algorithm proposed in this paper is 2.89, which
is conducted even if the doctor does not ask about other is 1.31 rounds less (31.19% decrease) compared with the
symptoms directly given the diagnosis. Subsequently, when average dialogue rounds in the dataset, and the improvement
the system asks the user for symptoms, the dialogue rounds of the disease prediction accuracy is 24.64%. When adding
is increased by 1. Our evaluation methods are 1) diagnostic constraints to the symptom screening of the disease diagnosis
accuracy, where the diagnosis succeeds when the predicted algorithm to limit the dialogue rounds to no more than the
result agrees with the true label, and 2) average dialogue rounds in the test set, the dialogue rounds of the disease
rounds, where the system gets the average dialogue rounds diagnosis algorithm is 2.65, and the accuracy of the disease
needed for each patient’s disease result. diagnosis algorithm is 85.91%, which is 6.34% lower than

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.
 IEEE Conference on ICT Solutions for eHealth (ICTS4eHealth 2023)

the initial constraint, but the overall accuracy of the disease

diagnosis algorithm is still 18.30% improvement.
In the DX dataset, the average dialogue rounds in the test
set was 2.76, and the average dialogue rounds for the disease
diagnosis algorithm in this paper was 3.04, an increase of
0.28 rounds over the average inquiring rounds in the dataset,
but a decrease of 0.32 (9.5%) over the inquiring rounds
in KR-DS. Adding constraints to the symptom screening
module of the disease diagnosis algorithm to limit it to no
more than the rounds in the dataset, the diagnostic accuracy
of the disease diagnosis algorithm was 90.38%, at which
point it improved by 12.50% and 16.38% compared to the
results of Bayesian-ex+im and KR-DS, respectively.
V. C ONCLUSION
In this paper, we propose disease diagnosis algorithm
based on Bayesian and a symptom screening algorithm
based on the difference of disease symptom set to solve
the problems of low diagnostic accuracy, long running time
and poor interpretability of existing reinforcement learning
models in this scenario. The algorithm achieved the best
results on two real-world datasets.
R EFERENCES
[1] Y. Hayashi, “A neural expert system with automated extraction of fuzzy
if-then rules and its application to medical diagnosis,” Advances in
neural information processing systems, vol. 3, 1990.
[2] Z. Wei, Q. Liu, B. Peng, H. Tou, T. Chen, X.-J. Huang, K.-F. Wong,
and X. Dai, “Task-oriented dialogue system for automatic diagnosis,”
in Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 2: Short Papers), 2018, pp. 201–
207.
[3] L. Xu, Q. Zhou, K. Gong, X. Liang, J. Tang, and L. Lin, “End-to-end
knowledge-routed relational dialogue system for automatic diagnosis,”
in Proceedings of the AAAI Conference on Artificial Intelligence,
vol. 33, no. 01, 2019, pp. 7346–7353.
[4] F. Nan, J. Wang, and V. Saligrama, “Feature-budgeted random forest,”
in International conference on machine learning. PMLR, 2015, pp.
1983–1991.
[5] D. Zakim, N. Braun, P. Fritz, and M. D. Alscher, “Underutilization of
information and knowledge in everyday medical practice: Evaluation
of a computer-based solution,” BMC Medical Informatics and Decision
Making, vol. 8, no. 1, pp. 1–12, 2008.
[6] K. Early, S. E. Fienberg, and J. Mankoff, “Test time feature ordering
with focus: Interactive predictions with minimal user burden,” in
Proceedings of the 2016 ACM International Joint Conference on
Pervasive and Ubiquitous Computing, 2016, pp. 992–1003.
[7] M. Kachuee, S. Darabi, B. Moatamed, and M. Sarrafzadeh, “Dynamic
feature acquisition using denoising autoencoders,” IEEE transactions
on neural networks and learning systems, vol. 30, no. 8, pp. 2252–
2262, 2018.
[8] Z. Wang, R. Wen, X. Chen, S. Cao, S.-L. Huang, B. Qian, and
Y. Zheng, “Online disease diagnosis with inductive heterogeneous
graph convolutional networks,” in Proceedings of the Web Conference
2021, 2021, pp. 3349–3358.
[9] A. Nesterov, B. Ibragimov, D. Umerenkov, A. Shelmanov, G. Zubkova,
and V. Kokh, “Neuralsympcheck: A symptom checking and disease
diagnostic neural model with logic regularization,” in Artificial In-
telligence in Medicine: 20th International Conference on Artificial
Intelligence in Medicine, AIME 2022, Halifax, NS, Canada, June 14–
17, 2022, Proceedings. Springer, 2022, pp. 76–87.
[10] A. F. Tchango, R. Goel, J. Martel, Z. Wen, G. M. Caron, and J. Ghosn,
“Towards trustworthy automatic diagnosis systems by emulating doc-
tors’ reasoning with deep reinforcement learning,” arXiv preprint
arXiv:2210.07198, 2022.

Authorized licensed use limited to: Zhejiang University. Downloaded on December 12,2024 at 13:01:50 UTC from IEEE Xplore. Restrictions apply.