text
stringlengths
0
89.3k
GLEU AccuracyZero LLMProfessional Safe
HelpfulPairwise
DISCMedLLM 37 MTQA MCQA Accuracy Zero Few LLMProactive
Professional HelpfulIndividual
Zhongjing 38 STQA MTQAHuman
LLMProfessional Safe
FluentPairwise
Case Study
BianQue 39 MTQABLEU ROUGE
PQAZero Human Case Study
Alpacare 40 STQA MTQA Accuracy Rouge Zero FewHuman
LLMProfessional HelpfulPairwise
Case Study
QilinMed 41 STQA MTQAF1Score BLEU
ROUGEZero Few SFT Human Case Study
Taiyi 42 MTQA NLP F1Score Accuracy Zero Human Case Study
ChiMedGPT 43STQA MTQA
MCQA NLPF1Scores Accuracy
BLEU ROUGEZero Few Human Case Study
MediTron 44 MCQA Accuracy Zero Few SFT
HuatuoGPTII 45STQA MTQA
MCQAAccuracy ZeroHuman
LLMProfessional Helpful
Patientfriendly FluentPairwise
Case Study
AntGLMMed 46 MCQA Accuracy Few SFT
GPTdoctor 47 STQAHuman
LLMProfessional Fluent
Patientfriendly
DoctorlikeIndividual
Pairwise
BioMistral 48 MCQA Accuracy Zero Few SFT
MMedLM 49 STQA MCQAAccuracy BLEU
ROUGE BERTScoreZero SFTHuman
LLMProfessional Pairwise
InMDX 9 Human Case Study
MeLLaMA 50 MCQA NLPAccuracy F1score
ROUGE BERTScoreZero Few SFT
JMLR 51 Human Case Study
BiMediX 52 MCQA Accuracy Zero
OncoGPT 53 STQA BERTScore Zero Human Case Study
Apollo 54 MCQA Accuracy Few
Qibo 55STQA MCQA
NLPAccuracy Rouge ZeroHuman
LLMProfessional Safe
FluentPairwiseJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 15
QA benchmark involves consecutive questions and answers
such as a dialog between a doctor and a patient This bench
mark assesses the ability of medical LLMs to perform medical
diagnosis understand context ensuring answers are based on
previous dialog and engage in interactive QA testing the
models ability to interact in realworld medical scenarios and
maintain fluent dialog The multiplechoice QA benchmark
presents a question along with multiple options requiring the
model to identify the correct answer With explicit correct
answers this benchmark allows precise quantification of the
medical LLMs capability Additionally some works not only
evaluate the options outputted by the model but also assess
reasoning content The varied multiplechoice questions eval
uate the models depth of knowledge across diverse medical
domains and its logical reasoning ability The commonly used
evaluation metrics for natural language generation tasks are
Accuracy BLEU 130 ROUGE 131 Distinct 132 and
GLEU 133 In addition to traditional evaluation metrics
certain studies have introduced evaluation metrics tailored to
healthcare dialog contexts For instance BianQue 39 pro
poses a novel evaluation metric named Proactive Questioning
Ability PAQ to assess the capability of LLMs in proactively
asking and querying questions to users
The majority of medical QA benchmarks comprise existing
medical QA datasets with only a limited number of bench
marks specifically gathered for particular evaluations Note
that the vast datasets employed in training LLMs may result
in overlap between training and test sets Therefore data de
duplication and data leakage checks are necessary to ensure
fairness of the test results
2 Protocols Medical LLMs are commonly assessed
through three protocols Zeroshot Fewshot and Task
specific Supervised Finetuning aiming to investigate model
performance across varying levels of contextual information
The zeroshot protocol is to evaluate the ability of a model to
directly perform a medical task with an unknown distribution
while assessing its generalization ability and adaptability It
only requires a natural language description of the downstream
task without providing specific examples to ask the model to
understand the medical task and use the intrinsic knowledge
to accomplish this task In the natural language understanding
task MeLLaMA 50 and Taiyi 42 designed prompt tem
plates for each downstream task to guide the output format of
the model to meet the expected requirements For the prompt
for QA tasks it usually give a role to the LLM and then guide
the models output style in line with expectations For instance
a prompt might be You are an experienced doctor please
respond to the patients inquiries with patience and kindness
Patient inputs Doctor Moreover addressing
medical issues often necessitates intricate multistep reasoning
Hence certain studies have employed prompt strategies to
enhance model performance For instance MediTron 44 uti
lizes the prompt strategy Zeroshot ChainofThought 134 to
guide the model towards a more comprehensive understanding