tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
consultation or the physicians clinical instructions to achieve
the expectations of users
Given the critical importance of safety in the medical field
this dimension must be assessed by human experts Other
dimensions can be evaluated by LLMs thereby increasing ef
ficiency For nonprofessional users the evaluation dimensions
focus on Patientfriendly and Fluency These two dimensions
match the users perspective and can measure the real feedback
of the model from users in real scenarios
3 Protocols Humancentric evaluation comprises two pri
mary protocols individual evaluation and pairwise evaluation
Individual evaluation involves a human evaluator or a LLM
scoring the output based on predefined dimensions with only
a single question and its corresponding output visible The
scoring range is a predetermined interval and the average
score for each dimension is statistically calculated as the
final result This metric quantitatively reflects the modelsoverall performance in each dimension identifying strengths
and weaknesses and guiding future optimizations on weaker
dimensions Pairwise evaluation involves a human evaluator
or a LLM comparing the outputs of two models based on
predefined dimensions and selecting the betterperforming
one Evaluation metrics include win rate the proportion of
evaluated models outperforming the comparison model tie
rate the proportion of the two that are equivalent and loss
rate the proportion of evaluated models underperforming the
comparison model In addition to minimize the positional
bias existing in the LLM as a judge the positions of the
two model outputs are exchanged for another comparison
Pairwise evaluation facilitates the rapid identification of the
bestperforming model from multiple candidates serving as a
foundation for model optimization and deployment
In addition to the above evaluation protocols some open
source projects use Case Study which relies entirely on
subjective evaluation It directly compares the responses of
multiple medical LLMs to the same question explicitly
demonstrating the differences in medical knowledge and re
sponse style between them It does not establish metrics
for measuring model performance relying entirely on the
subjective perception of users
V C HALLENGES AND FUTURE DIRECTIONS
We analyze the current challenges of medical LLMs from
the perspectives of data methodology and evaluation and
propose corresponding future research directions
A Data Privacy
To enable medical LLMs to achieve capabilities comparable
to those of medical professionals it is essential to integrate
more real medical data into these models to enhance their med
ical knowledge Data from professional medical organizations
tend to be more authentic and reliable than data sourced from
the web However due to the sensitive nature of medical data
and associated policies such data is seldom publicly available
Consequently a significant challenge is to ensure the secure
utilization of medical data within organizations while
safeguarding patient privacy One approach to addressing
this issue is federated learning 137 a distributed training
method that enables model parameters to be shared across
multiple organizations without sharing the data itself By
keeping data stored locally federated learning preserves data
privacy while allowing the model to learn from diverse data
sources thus enhancing its generalization and performance
Federated learning has been applied to medical images 138
sensor data 139 and natural language text 140 promoting
the secure use of health data among medical organizations
Currently federated learning has several shortcomings in
the medical field that require further investigation Differential
privacy 141 can be integrated with federated learning to
enhance data privacy protection but this integration degrades
model performance How to balance privacy protection and
model performance is a concern Additionally the quality
and quantity of data varies among healthcare organizations
and how to evaluate the contribution of each organizationJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 17
is a key topic Finally medical LLMs have a significantly
higher number of parameters compared to pretrained language
models raising the issue of optimizing communication and
computational overhead during the transmission and aggrega
tion of parameters in largescale models
B Knowledge Augmentation
Before providing specific diagnoses and prescriptions
physicians typically inquire about patients medical and allergy
histories However the capabilities of medical LLMs rely
mainly on training data that do not contain information about
the currently attending patients Consequently proposed diag
noses and prescriptions may not be applicable to individual
patients Therefore a significant challenge is enabling med
ical LLMs to provide personalized services to patients or
doctors using realtime information RetrievalAugmented
Generation RAG 142 presents a feasible solution This
method first retrieves content relevant to the user query from
external realtime sources eg medical news personal elec
tronic medical records It then integrates the user query with
the retrieved content into a coherent prompt and the model
subsequently responds based on its own knowledge and the
prompt Several medical LLMs work have been developed
to enhance responses using external knowledge thereby im
proving performance and mitigating hallucination problems
For instance ChatDoctor 25 employs online resources and
custom offline medical databases as an external knowledge
base to enhance accuracy in medical tasks through RAG
And JMLR 51 trains medical LLMs together with RAG
to improve the models ability to process medical knowledge

consultation or the physicians clinical instructions to achieve

the expectations of users

Given the critical importance of safety in the medical field

this dimension must be assessed by human experts Other

dimensions can be evaluated by LLMs thereby increasing ef

ficiency For nonprofessional users the evaluation dimensions

focus on Patientfriendly and Fluency These two dimensions

match the users perspective and can measure the real feedback

of the model from users in real scenarios

3 Protocols Humancentric evaluation comprises two pri

mary protocols individual evaluation and pairwise evaluation

Individual evaluation involves a human evaluator or a LLM

scoring the output based on predefined dimensions with only

a single question and its corresponding output visible The

scoring range is a predetermined interval and the average

score for each dimension is statistically calculated as the

final result This metric quantitatively reflects the modelsoverall performance in each dimension identifying strengths

and weaknesses and guiding future optimizations on weaker

dimensions Pairwise evaluation involves a human evaluator

or a LLM comparing the outputs of two models based on

predefined dimensions and selecting the betterperforming

one Evaluation metrics include win rate the proportion of

evaluated models outperforming the comparison model tie

rate the proportion of the two that are equivalent and loss

rate the proportion of evaluated models underperforming the

comparison model In addition to minimize the positional

bias existing in the LLM as a judge the positions of the

two model outputs are exchanged for another comparison

Pairwise evaluation facilitates the rapid identification of the

bestperforming model from multiple candidates serving as a

foundation for model optimization and deployment

In addition to the above evaluation protocols some open

source projects use Case Study which relies entirely on

subjective evaluation It directly compares the responses of

multiple medical LLMs to the same question explicitly

demonstrating the differences in medical knowledge and re

sponse style between them It does not establish metrics

for measuring model performance relying entirely on the

subjective perception of users

V C HALLENGES AND FUTURE DIRECTIONS

We analyze the current challenges of medical LLMs from

the perspectives of data methodology and evaluation and

propose corresponding future research directions

A Data Privacy

To enable medical LLMs to achieve capabilities comparable

to those of medical professionals it is essential to integrate

more real medical data into these models to enhance their med

ical knowledge Data from professional medical organizations

tend to be more authentic and reliable than data sourced from

the web However due to the sensitive nature of medical data

and associated policies such data is seldom publicly available

Consequently a significant challenge is to ensure the secure

utilization of medical data within organizations while

safeguarding patient privacy One approach to addressing

this issue is federated learning 137 a distributed training

method that enables model parameters to be shared across

multiple organizations without sharing the data itself By

keeping data stored locally federated learning preserves data

privacy while allowing the model to learn from diverse data

sources thus enhancing its generalization and performance

Federated learning has been applied to medical images 138

sensor data 139 and natural language text 140 promoting

the secure use of health data among medical organizations

Currently federated learning has several shortcomings in

the medical field that require further investigation Differential

privacy 141 can be integrated with federated learning to

enhance data privacy protection but this integration degrades

model performance How to balance privacy protection and

model performance is a concern Additionally the quality

and quantity of data varies among healthcare organizations

and how to evaluate the contribution of each organizationJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 17

is a key topic Finally medical LLMs have a significantly

higher number of parameters compared to pretrained language

models raising the issue of optimizing communication and

computational overhead during the transmission and aggrega

tion of parameters in largescale models

B Knowledge Augmentation

Before providing specific diagnoses and prescriptions

physicians typically inquire about patients medical and allergy

histories However the capabilities of medical LLMs rely

mainly on training data that do not contain information about

the currently attending patients Consequently proposed diag

noses and prescriptions may not be applicable to individual

patients Therefore a significant challenge is enabling med

ical LLMs to provide personalized services to patients or

doctors using realtime information RetrievalAugmented

Generation RAG 142 presents a feasible solution This

method first retrieves content relevant to the user query from

external realtime sources eg medical news personal elec

tronic medical records It then integrates the user query with

the retrieved content into a coherent prompt and the model

subsequently responds based on its own knowledge and the

prompt Several medical LLMs work have been developed

to enhance responses using external knowledge thereby im

proving performance and mitigating hallucination problems

For instance ChatDoctor 25 employs online resources and

custom offline medical databases as an external knowledge

base to enhance accuracy in medical tasks through RAG

And JMLR 51 trains medical LLMs together with RAG

to improve the models ability to process medical knowledge