Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

Applied Mathematics and Nonlinear Sciences

https://www.sciendo.com

Development Status and Strategy Analysis of Medical Big Models

Guozheng Zhang1, Jinxin Liu2,†, Jian Mu1

1. China Mobile (Chengdu) Information and Communication Technology Co., Ltd., Chengdu,
Sichuan, 610000, China.
2. China Mobile Communications Group Co., Ltd., Beijing, Beijing, 100000, China.

Submission Info

Communicated by Z. Sabir
Received January 2, 2024
Accepted January 7, 2024
Available online February 26, 2024

Abstract
It is imperative to embark on a significant model-driven technology route for the intelligent development of the medical
industry. This paper constructs a medical big model through three aspects: medical drug recommendation, medical data
sampling, and medical image segmentation. The link between symptoms and drugs is established through the PALAS
algorithm, the unbalanced numerical dataset is solved by using the oversampling SMOTE method, and the source domain
of medical images is labeled by the MCDIFL method to adapt to the unsupervised domain in medical image segmentation.
Finally, the development trend of medical macromodeling is discussed, and the data of diabetic patients admitted to
Hospital X is used as a study case to specifically explore the effect of medical macromodeling in healthcare. The results
show that the data of diabetic patient A was inputted into the medical extensive model analysis to obtain that the average
values of patient A's blood glucose value in the first three years were 7.13, 9.34, and 7.06 mmol/L, respectively, which
experienced the evolution from mild to high and then to soft. The results can help medical personnel to make a scientific
treatment plan for the patient. This study promotes the innovative application and development of artificial intelligence
technology in medical services.

Keywords: Data sampling; Image segmentation; Medical macromodeling; PALAS algorithm; Oversampling
SMOTE.
AMS 2010 codes: 97M80

†Corresponding author.
Email address: [email protected] ISSN 2444-8656
https://doi.org/10.2478/amns-2024-0325
© 2023 Guozheng Zhang, Jinxin Liu and Jian Mu, published by Sciendo.
This work is licensed under the Creative Commons Attribution alone 4.0 License.
 2 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

1 Introduction

With the rapid development of AI technology, the field of healthcare services has been affected and
changed as never before. Advanced AI technologies offer great potential for improving healthcare,
improving patient experience, and promoting medical research [1]. In the medical field, generative
language models represented by ChatGPT can better accomplish medical assistant-type work, such
as booking appointments, information management, and health counseling, improving the efficiency
and quality of medical staff and lowering the threshold of patient information access [2-3]. The large
model can also give diagnostic suggestions, treatment plans, and class case recommendations on the
condition, thus opening up the doctor's diagnosis and treatment ideas, suggesting the risk of
medication and clinical guidelines, and realizing clinically assisted decision-making [4-5].

At the same time, it helps medical professionals to read a large amount of literature, refine core ideas
and conclusions, facilitate their understanding of industry dynamics, accelerate the experimental
process, etc. [6]. With the improvement of the knowledge system and the enhancement of cognitive
ability, the big model can complete the diagnosis of conventional diseases or medical advice with the
help of question and answer interaction form, search and recommend similar cases, so as to be able
to perform some of the work of general practitioners [7-9]. With the deepening of the application of
big models in the medical industry, the time and energy of doctors are expected to be further liberated,
thus alleviating the problem of scarce resources of medical experts.

Large models are viewed in many contexts as a class of fundamental models, i.e., models that are
trained on large-scale data and can be adapted to a wide range of downstream tasks. Liu, F. et al.
explored the application of Medical Multimodal Large Language Models (Med-MLLMs) in
pandemics. The Med-MLLM large models were able to recognize massive amounts of medical data
across visual modalities and textual modalities and, from these data, to identify rare pandemic
diseases and thus make rapid deployments [10]. Shah, N. H et al. explored the application of Large
Language Models (LLMs) in the medical field and found that LLMs perform well in following
instructions and generating human-like replies but tend to be ineffective in specific specialized
medical domains through the training of the models, the models have been increasingly used to among
performing medical tasks [11].

Chirigati F found that in treating patients, physicians often need to consider a large amount of data,
including imaging reports, case reports, physical examination reports, and other data, which is a
process and cumbersome, so he proposed to input data from electronic health records (EHRs) into
clinical prediction models [12]. Zhang, T et al. found that pre-trained language models (PLMs)
applied in the medical domain would make it difficult to understand the specific content of the text
due to the medical terminology and its complex relationships, and therefore proposed a solution to
extract information about heterogeneous entities using deep structured semantics [13]. Murphy, C
and Thomas, F. P. believe that ChatGPT, a neural network model developed by OpenAI and trained
on large datasets, will revolutionize the medical field, with ChatGPT being able to generate
contextually relevant and informative texts of similar studies, speeding up the process of data
refinement for medical researchers [14].

Suganya, A., and Aarthy, S. L. believe that the difficulty in the treatment of Alzheimer's and
Parkinson's disease, which are highly prevalent neurodegenerative diseases, lies in early diagnosis
because the diagnosis of neurodegenerative diseases requires extremely professional medical
personnel to make a comprehensive judgment on the patient's physical examination indicators and
historical medical history, so the authors propose to apply deep Xi (DL) to the detection of these
diseases [15]. Taghizadeh, N. et al. constructed a Sina-BERT model applied to the online medical
field through the pre-training of a large-scale medical information corpus, which is able to accurately
 Development Status and Strategy Analysis of Medical Big Models 3

retrieve and categorize all kinds of medical problems, and the medical staff can retrieve past cases
and make a diagnosis quickly through this model [16].

In this paper, we first learn the potential symptom space through the PALAS algorithm, establish the
connection between symptoms and prescription drugs, train PALAS using the alternating
optimization algorithm, and then construct the drug recommendation model. The Euclidean distance
between a few classes of samples is analyzed by the SMOTE method to solve the imbalance
phenomenon of the numerical dataset and construct a medical data sampling model. A medical image
segmentation model is constructed based on the multiple constraints domain invariant feature learning
(MCDIFL) method to solve the problem of unsupervised domain adaptation in medical image
segmentation. Secondly, through the medical 5G terminal information processing technology, the data
acquisition processing transmission framework among the hospital multi-source heterogeneous
medical equipment, 5G terminal, and telemedicine center is constructed, which solves the needs of
the medical system in various aspects, such as rapid data acquisition, processing, and transmission.
Finally, the development trend of the medical big model is discussed from three aspects of: policy,
economic, and technical environment, and the specific applications of the medical big model are
analyzed.

2 Method

2.1 Data Analytics Techniques Based on Healthcare Megamodels

2.1.1 PALAS-based drug recommendation models

PALAS aims to learn the latent symptom space to better model the relationship between observed
symptoms and prescribed medications. The novel model structure of the Prescribing System for
Latent Symptom Learning (PALAS) is shown in Figure 1. The model is inspired by the following
observations:

1) Drugs are sparsely labeled, and a drug relieves only a small fraction of all symptoms. Similarly,
few patients are prescribed a variety of therapeutic drugs. Instead, most patients will choose
only a very small number of all candidate therapeutic drugs, resulting in sparse drug labeling
for each subject.

2) Large semantic differences between observed symptoms and therapeutic drugs require the
specialized experience of clinicians to understand. Therefore, the relationship between
symptoms and drug labels should not be modeled directly, as this may limit the performance
of the predictive model.
 4 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

Multimodal Latent
Drug labelling
representation symptom space

Motor Characteristic-based representation

symptoms

Similarity-based representation

Non-motor
symptoms Feature-based representation

Figure 1. Novel model structure of prescription system for Latent symptom learning (PALAS)

The PALAS model addresses the above problem by learning a latent symptom space that exploits the
link between latent symptoms and medications. Specifically, assume that given a multimodal
representation of all samples X i  ( n+m)di ( i = 1,,6) and a matrix of true drug labels for the n
n c
training samples Y  B . where B = {0,1} , c are the total number of drugs, the goal is to predict
the drug labels for the m test samples. In order to keep the dimensionality consistent,
Y  B( n + m )c = [Y ;0] is introduced as an additional drug class labeling matrix, where the Y last m
columns are set to 0 due to the unavailability of the test samples during the training phase.

The PALAS model is proposed to accomplish this prediction task, in which a potential symptom
representation P  ( n+ m)k is learned to correlate observed symptoms with prescribed medications,
where k denotes the number of potential symptoms.The main objective function of PALAS is
mathematically defined as follows:

( )
s
arg min F (U i , P, V ) =  X iU i − P +  Ui
2 2
F F
U i , P ,V i =1 (1)
+ J (Y − PV ) + V 1,
2
F

di  k
Where U i  is the observation-to-potential-symptom transformation matrix to ensure that the
k c
potential symptom space P can represent the i rd modality, i.e., X i . V  is the symptom-
to-drug transformation matrix used to associate different potential symptoms with different drugs. V
is restricted to be sparse because typical drugs only relieve very few symptoms. Also, from a clinical
( n + m )( n + m )
point of view, very few patients take all possible medications. J  B A binary diagonal
matrix can be expressed as:

 n m

J = diag 1,,1,10,, 0  (2)
 
 

It is used to represent the training samples that are in the direct-push setting so that only the training
samples will participate in predicting the loss. In this way, although the test samples do not help in
 Development Status and Strategy Analysis of Medical Big Models 5

drug label prediction, combining the test samples can help in learning P and U i , thus solving the
problem of limited samples in this study. Where  and  are regularization parameters used to
equalize the different terms in the equation.

2.1.2 SMOTE-based model for sampling healthcare data

Medical data comes from various types of medical devices; due to the non-homogeneous nature of
the data, which leads to the input to the Analysis of medical macromodeling, the dataset has
unbalanced linearity. Commonly used are Synthetic Minority Oversampling Technique (SMOTE),
downsampling, and Ensemble methods. In this paper, the oversampling SMOTE method will be used
to solve the imbalance of the numerical dataset.

The oversampling SMOTE method improves the problem of model overfitting in random
oversampling by simply replicating a few classes of samples. The basic process is by analyzing the
Euclidean distance between the few classes of samples. The minority class samples are synthesized
between the Euclidean distances of the minority class samples so that their Euclidean distances from
other minority class samples satisfy the analytical law, and then the synthesized samples are added to
the dataset. The specific process of synthesizing the minority class samples by the SMOTE algorithm
is illustrated in Fig. 2, where the blue squares indicate the minority class samples and the orange
circles indicate the majority class samples. The formula for the location of the synthesized minority
class sample is:

xnew = xi + rand (0,1)  xi − xi (3)

f2

yi

Generate
xi sample

f1

Figure 2. SMOTE method synthesizes a few class samples

2.1.3 MCDIFL-based medical image segmentation model

In medical macromodeling, this paper proposes the Multiple Constraint-based Domain Invariant
Feature Learning (MCDIFL) method and uses it to solve the problem of unsupervised domain
 6 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

adaptation in medical image segmentation. Specifically, given a source domain dataset with N S
labeled image as:

( x , y )
NS
XS = s
i
s
i
(4)
i =1

where, for each medical grayscale image slice xis  H W 1 , there are its corresponding height of H ,
height and width of W , and C channel annotations corresponding to pixel-level class labels
 
NT
yis  {0,1}H W C , and NT unlabeled image target domain datasets X T = xit .
i =1

The methodological framework proposed in this paper consists of a total of seven networks with
shared encoder E ( x;enc ) , pixel-level classification task network, i.e., segmentation network
T ( z;t ) , decoder that reconstructs the source domain image from hidden variables z , decoder that
reconstructs the target domain image from hidden variables U S z;us( ) and decoder that
reconstructs the target domain image from hidden variables U ( z; ) , semantic spatial T ut

discriminator D ( p; ) , source domain image discriminator D ( x; ) , and target domain image
dp S ds

discriminator D ( x; ) , respectively, where  ,  ,  ,  ,  ,  , and  denote the

T dt enc t us ut dp ds dt

learning parameters in each of the networks. learning parameters in each network.

Given a source domain image xis as input, encoder E outputs hidden variable features z s to
represent deep feature high-level semantics to characterize domain invariant properties. Their
computational formulas are respectively:

(
zis = E xis ; c ) (5)

(
yˆis = T zis ;t ) (6)

(
xˆis →s = U S zis ;us ) (7)

(
xˆis →t = UT zis ;ut ) (8)

Similarly, the target domain image xit may be similarly encoded and decoded. Specifically:

(
zit = E xit ; c ) (9)

(
yˆit = T zit ; t ) (10)

(
xˆit →s = U S zit ;us ) (11)

(
xˆi→t = UT zit ;ut ) (12)
 Development Status and Strategy Analysis of Medical Big Models 7

Where, yˆi  {0,1}H W C denotes the pixel level classification prediction results obtained by the
segmentation network. x̂ Denotes the generated image, after this network four kinds of generated
s→s
images can be obtained, which are source domain reconstructed image xˆ , target domain
t →t s →t
reconstructed image xˆ , source domain transformed into target domain image xˆ , target domain
t→s
transformed into source domain image xˆ These outputs are used for the discriminator to
determine the truth and falsehood, and the loss is computed by the constraints at each level, so that
z s and z t and yˆ s and yˆ t are as similar as possible and are feature representations that are in the
same space, i.e., the constraints are respectively:

( ) (
P z s | x s = P z t | xt ) (13)

( ) (
P y s | x s = P y t | xt ) (14)

Such a framework optimizes the following loss functions and performs end-to-end training:

( )
L = seg Lseg + rec Lrec + I LSI + LTI + P LP (15)

2.2 Medical 5G terminal information processing technology

2.2.1 Monitoring of the operational status of medical equipment

The 5G terminal display module includes the data reception for the sensing devices, as well as the
Monitoring of the status of the 5G terminal devices, and also the GPS positioning information of the
5G terminal. In the 5G terminal human-computer interaction display function module for mobile
hospital medical applications, the functional program flow of the status monitoring process is shown
in Figure 3. After the initialization of the system, the 5G terminal applications run individually to
obtain the monitored information in real-time. Transmitted to the component of page display through
the shared memory, the display module will process the received data and display it on the page; at
the same time, the background real-time monitoring of the port to determine whether there is any data
to realize the dynamic page update.
 8 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

Start

System initialisation

Get terminal location Detect terminal hardware

Detect peripheral devices
information system operation status

Network
System resource Peripheral port Device real-time
communication
consumption status data
status

Sharing data information between terminal and host computer

Each page display component obtains data and displays it

Page content scrolling refresh

Figure 3. Medical status monitoring process function program flow

The whole page display module follows the humanized page design, and by partitioning different
information, it is easier for healthcare professionals to obtain key information accurately and in real-
time. In the terminal display module, the upload and download rates of the 5G network will be
displayed in real-time, the medical equipment area will display the collected medical data and whether
the peripheral equipment is well connected in real-time, and there is also an area for displaying the
decision-making opinions of telemedicine experts.

2.2.2 5G Medical Data Acquisition Module

This paper investigates the data specifications and applicable scenarios of medical equipment from
multiple manufacturers in the industry, analyzes the interface heterogeneity of 5G terminals on the
medical equipment side, and designs and implements real-time acquisition, processing and display,
and information integration and transmission of multi-source data from medical equipment with
heterogeneous interfaces in mobile hospitals, so as to transform the underlying medical data into the
high-quality dataset required by upper-layer applications, and to satisfy the needs of mobile medical
systems in various aspects such as data acquisition, processing, and transmission. It transforms the
underlying medical data into high-quality data sets required by the upper-layer applications and meets
the needs of mobile medical systems in terms of data acquisition, processing, and transmission.

In view of the heterogeneous nature of medical equipment interfaces, epoll I/O multiplexing
technology is used to realize the non-blocking reception of heterogeneous medical equipment data
from multiple sources, and the overall process of multi-threaded acquisition of heterogeneous data
from 5G industry terminals is shown in Figure 4. Different interfaces of medical devices collect data
corresponding to different threads, and each of them runs independently. The 5G terminal sends
 Development Status and Strategy Analysis of Medical Big Models 9

control commands and receives relevant medical data to realize communication between the 5G
terminal and different devices. It also uses shared memory to realize real-time storage and invocation
of cached data. 5G terminal acts as an embedded middleware in the whole process and uniformly
collects medical data into the terminal to facilitate the Analysis and processing of the medical model
in the next step. Step of parsing and processing of the medical big model.

Installation and configuration of heterogeneous

interface drivers

Connect the corresponding physical interface and run

the system programme

Configure programs according to different physical

interfaces to receive data streams and process them.

Read available device files

Terminal server application Enable Bluetooth/USB

Read terminal serial port files Read camera devices
listens devices

Configure parameters (baud Receive client requests, read Configure parameters such as
Data storage
rate, parity code, etc.) network device files ton rate

Send commands to open the Parsing according to

Confirms connection Capture video streams
port corresponding protocols

Read data, write to memory Send startup data Close device file Cache and forward

Determination of packet size,

Store data, parse values End reading
order and other parameters

Successful connection, data

Close serial port
reception

Storage parsing

Closed request/listening

Closed data reception and acquisition procedure

Figure 4. Data collection process of heterogeneous interface devices on 5G terminals

 10 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

3 Results and discussion

3.1 Development trend analysis of medical macromodeling

1) Policy environment analysis

In March 2020, in the Outline of the National Healthcare Service System Plan (2020-2025), it was
proposed to carry out the planning of cloud services for a healthy China and actively apply new
technologies such as cloud computing, big data, Internet of Things, mobile Internet, 5G technology,
AI big models, etc., to promote health information services and smart medical services benefiting the
whole population. Actively promote the development of mobile Internet and telemedicine services.
Realize information sharing and business synergy of medical services, medical insurance, and public
health services at all levels.

2) Economic Environment Analysis

With the technical characteristics and strong network coverage capability of 5G, the development of
mobile, intelligent, and remote medical applications can alleviate various blockages and bottlenecks
in the development of medical business. Since the pilot application of 5G with various medical and
healthcare scenarios began in 2019, the combination of 5G and medical and healthcare has entered a
large-scale construction stage from pilots, and hospitals and other medical institutions have begun to
try to deeply integrate 5G with more medical business scenarios and apply it practically.

Driven by the new medical reform policy, the cost of building a large data model in China's medical
and health field has been expanding, and the market for various products and services has grown
rapidly. Figure 5 shows the statistical results of CCWResearch's “2022 Research Report on Digital
Construction and AI Large Model Application Trends in China's Medical Industry”. The data shows
that in 2022, the scale of AI model investment in the medical and health industry will reach 31.812
billion yuan, a year-on-year increase of 0.42 compared with 2021, continuing to show a trend of
steady and rapid growth, and the total investment will increase by 21.45 times compared with 2012.
With this continuous growth trend, it is expected that the total investment in large AI models in the
medical and health industry will reach 36.15 billion yuan in 2023.

Figure 5. Investment scale of large data model construction in China's medical field

3) Analysis of the technological environment

On December 1, 2022, Open AI, a U.S. artificial intelligence company, launched ChatGPT, a big
language model, which reached 100 million active users about 2 months later, making it the fastest-
 Development Status and Strategy Analysis of Medical Big Models 11

growing consumer application in history.ChatGPT, with its powerful natural language processing and
conversational interaction features, has demonstrated a wide range of applications in many fields.
Currently, big language models such as ChatGPT have begun to try to be applied in medical services
such as case writing, report generation, case summarization, and assisted decision-making [17]. In
the future, it may be further applied to medical scenarios such as case records, medical consultation,
popularization education, patient management, academic communication, etc. The feasibility of the
application of AI technology in the field of healthcare services has been confirmed, and the future
prospects are promising.

3.2 Analysis of the application of healthcare big models in healthcare

3.2.1 Patient case studies

The data utilized in this paper is a data set of diabetic patients admitted to Hospital X, which results
in abnormal data. There are 13,182 diabetes data identified by correlating it with the medical macro
model. The data taken is a total of 26 medical data that have an impact on diabetes. It contains 11
physiological information such as body temperature, pulse rate, systolic blood pressure, diastolic
blood pressure, height, weight, hip circumference, and body mass index (BMI). And 15 blood
chemistry information such as serum creatinine, total bilirubin, conjugated bilirubin, and total
cholesterol.

Its structured data mainly contains seven attributes, i.e., procedure ID, name, gender, age, weight,
height, and department. The differences in age, height, and weight by gender are shown in Figures 6-
8, where it can be seen that the female: male ratio was 56.58:43.42 in the whole data. It was also
noted that there were significant differences in weight and height between males and females, where
female patients were mostly concentrated in the range of 50kg to 60kg, while male patients were
mostly concentrated in the range of 60kg to 70kg. The mean weights of male and female patients
were 64.33 and 53.12 kg, respectively, and the heights of female patients ranged from 145 cm to 165
cm, while that of male patients ranged from 160 cm to 175 cm. In terms of age, most of the patients
were between 23 and 68 years old, but the average age of the female patients was 39 years old, and
the proportion of young female patients was higher than that of male patients.

Figure 6. Weight difference between male and Figure 7. Height difference between male and
female patients female patients
 12 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

Figure 8. Age difference between male and female patients

3.2.2 Analysis of the evolution of patient indicators

In this paper, after processing the data of diabetic patients in Hospital X through the previous
subsection, we explored the prediction of glycemic changes in diabetic patients with lifestyle habit
data using the healthcare macro model constructed in this paper. The value of fasting blood glucose
value change over the time span was used as a predictive label, and three basic factors, basal blood
glucose value, height, and weight, were used as predictors, along with five lifestyle habit factors,
namely, smoking, staple food, daily exercise time, weekly exercise frequency, and alcohol
consumption. The reference values of fasting blood glucose for diabetic patients were mild diabetes
(6.8-8.5 mmol/L), moderate diabetes (8.5-10.8 mmol/L), and severe diabetes (≥10.8 mmol/L).
Accordingly, this article classifies fasting blood glucose values into four classes. Normal blood
glucose value is less than 7.0 mmol/L. Meanwhile, the year is used as the time unit for the evolution
of diabetes. The medical base data were imported into the medical grand model for Analysis, and the
following analysis results were obtained:

1) Blood glucose changes in diabetic patients are shown in Figure 9, and when a patient is
selected, the system highlights by default the nodes and migration lines for the subsequent
year associated with this blood glucose level node. For example, the 2019 blood glucose value
of patient A in the figure is mild glucose (7.13 mmol/L); in the second year, the blood glucose
value increases to moderate glucose (9.34 mmol/L), and in the third year, there is a remission
and returns to mild glucose (7.06 mmol/L). From this, it can help analysts to visualize the
change in blood glucose levels of people with this blood glucose level after one year's
influence of lifestyle habits, which can provide scientific decision-making for patients.
 Development Status and Strategy Analysis of Medical Big Models 13

2019 2020 2021 2022

< 6.8

6.8~8.5

8.5~10.8

≥ 10.8

Normal Mild blood Moderate Severe

blood sugar sugar blood sugar blood sugar

Figure 9. Trend of blood glucose value in diabetic patients

2) The lifestyle habits of diabetic patients are combined with the graph of glycemic evolution to
obtain the results of the ranking of the importance of the characteristics of the lifestyle habits
as shown in Figure 10, which displays the distribution of the lifestyle habits of this part of the
population with the evolution of the glycemic level, according to the glycemic level chosen
by the analyst. It can be seen that the magnitude of the basal blood glucose value is the most
important characteristic influencing the evolution of blood glucose, with an influence factor
of 0.352. The amount of staple food comes next, with an influence of 0.169 on the value of
blood glucose, followed by systolic blood pressure, diastolic blood pressure, body mass index,
time of exercise per day, smoking, number of weekly exercise sessions, and alcohol
consumption, in that order. Screening the population with elevated blood glucose and the
population with lowered blood glucose and comparing the differences in the distribution of
lifestyle habits between these two populations can be used to understand the effect of lifestyle
habits on the elevation and lowering of blood glucose in patients.

Figure 10. Ranking results of the importance of lifestyle characteristics.

 14 Guozheng Zhang, Jinxin Liu and Jian Mu. Applied Mathematics and Nonlinear Sciences, 9(1) (2024) 1-15

4 Conclusion

The existence of problems, such as the uneven distribution of medical resources and the disparity
between urban and rural medical service levels, has pushed many medical institutions towards
intelligence and informatization. In this paper, we constructed a grand medical model through PALAS,
sampling SMOTE, MCDIFL, and other methods and completed the empirical application of the
model. The main research of this paper is summarized as follows:

1) From the point of view of the development of the medical big model trend, the total investment
scale of the AI big model in the medical field reached 31.812 billion yuan in 2022, with a
year-on-year growth of 0.42, showing a steady and high growth trend. In the past 10 years,
benefiting from the development and application of information technology, the rapid
development of medical technology, and a new round of medical changes, the medical model
in ChatGPT technology boost, high investment in the economy, and government support is
showing “emergence” of the fierce development trend.

2) From the practical application of the medical big model, the model can accurately determine
the maximum influence factor of diabetic patients is the basal blood glucose value (0.352),
and the secondary influence factor is the amount of staple food (0.169), and the result can
make scientific treatment decisions for medical personnel.

In summary, the application of big medical models in medical services and medical research has great
potential and value; they can alleviate the problem of strained medical resources, improve the quality
and efficiency of medical care, reduce medical costs, and promote medical innovation and progress.

In the future, with the continuous accumulation of medical big data and the continuous progress of
AI big language modeling technology, the medical field will usher in a broader development space
and opportunities. It can be expected that in the near future, more accurate, efficient, and intelligent
medical diagnosis and treatment will be realized by combining medical big data with artificial
intelligence big models. It will bring great opportunities for medical research, public health
policymaking, medical service provision, and business model innovation and make greater
contributions to human health.

References
[1] Madden, M. G., Mcnicholas, B. A., & Laffey, J. G. (2023). Assessing the usefulness of a large language
model to query and summarize unstructured medical notes in intensive care. Intensive care medicine.
[2] Zhang, B. (2023). Chatgpt, an opportunity to understand more about language models. Medical Reference
Services Quarterly, 42(2), 194-201.
[3] Li, H., Moon, J., Purkayastha, S., Celi, L., Trivedi, H., & Gichoya, J. (2023). Ethics of large language
models in medicine and medical research. The Lancet. Digital health.
[4] Blanc, C., Bailly, A., Francis, L., Guillotin, T., Jamal, F., & Béchara.Wakim, et al. (2022). Flaubert vs
camembert: understanding patient's answers by a french medical chatbot. Artificial intelligence in
medicine, 127, 102264.
[5] Sam, A., Lam, G., & Amin, A. (2023). Medical exams in the era of accessible artificial intelligence.
Medical Teacher, 45, 669 - 670.
[6] Gasmi, K. (2022). Medical text classification based on an optimized machine learning and external
semantic resource. Journal of circuits, systems and computers.
[7] Ngo, G., Nguyen, M. L., Chen, N. F., & Sabuncu, M. (2022). A transformer-based neural language model
that synthesizes brain activation maps from free-form text queries. Medical image analysis, 81, 102540.
 Development Status and Strategy Analysis of Medical Big Models 15

[8] Lan, Y., He, S., Zeng, X., Liu, S., Liu, K., & Zhao, J. (2021). Path-based knowledge reasoning with textual
semantic information for medical knowledge graph completion.
[9] Jiang, L. Y., Liu, X. C., Nejatian, N. P., Nasir-Moin, M., & Al, D. W. E. (2023). Health system-scale
language models are all-purpose prediction engines. Nature.
[10] Liu, F., Zhu, T., Wu, X., Yang, B., You, C., & Wang, C., et al. (2023). A medical multimodal large
language model for future pandemics. npj Digital Medicine, 6(1).
[11] Shah, N. H., Entwistle, D., & Pfeffer, M. A. (2023). Creation and adoption of large language models in
medicine. JAMA: the Journal of the American Medical Association.
[12] Chirigati, F. (2023). A language model for medical predictive tasks. Nature Computational Science.
[13] Zhang, T., Cai, Z., Wang, C., Qiu, M., Yang, B., & He, X. (2021). Smedbert: a knowledge-enhanced pre-
trained language model with structured semantics for medical text mining.
[14] Murphy, C., & Thomas, F. P. (2023). Navigating the ai revolution: this journal's journey continues. The
Journal of Spinal Cord Medicine, 46, 529 - 530.
[15] Suganya, A., & Aarthy, S. L. (2023). Application of deep learning in the diagnosis of alzheimer's and
parkinson's disease-a review. Current medical imaging.
[16] Taghizadeh, N., Doostmohammadi, E., Seifossadat, E., Rabiee, H. R., & Tahaei, M. S. (2021). Sina-bert:
a pre-trained language model for Analysis of medical texts in persian.
[17] Kleesiek, J., Wu, Y., Tiglic, G., Egger, J., & Bian, J. (2023). An opinion on chatgpt in health care—
written by humans only. The Journal of Nuclear Medicine, 64, 701 - 703.