tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
sources Specifically medical consultation platforms allow pa
tients to communicate with doctors online for medical advice
diagnoses and treatment plans enhancing the convenience
and accessibility of healthcare services These platforms also
offer doctors a flexible way to provide care more efficiently
Common medical websites include Ding Xiang Yuan 81
Ping An Good Doctor 82 Teladoc Health 83 98point6
84 and Haodf 85 For training medical LLMs LLMs
conversation data from medical consultation platforms is par
ticularly valuable This data is generally more accessible to
the public compared to academic literature and textbooks due
to its realworld colloquial nature Additionally the multi
round questionandanswer format helps improve the models
ability to process contextual information Several models such
as MMedLM 49 ChatDoctor 25 and GPTDoctor 47 use
this type of data to align their response styles with those of real
doctors However data from medical consultation platforms
must be collected with strict adherence to privacy regulations
to ensure patient information security The quality of this
data can vary with some information being inaccurate or
incomplete Therefore it is essential to screen and clean the
data to maintain the accuracy and reliability of the medical
information used for training
Medical Guidelines also known as clinical or practice
guidelines are documents that provide recommendations for
medical decisionmaking including the diagnosis manage
ment and treatment of specific health conditions These
guidelines are critical for medical practice and are typically
developed by professional medical organizations government
agencies or academic groups 86 Common sources include
CCO 87 CDC 88 NICE 89 and WHO 90 Medical
guidelines offer practical information across a broad spec
trum of diseases and clinical scenarios aiding medical LLMs
in learning diagnostic and therapeutic decisionmaking The
standardized format and consensusdriven content of these
guidelines enable models to assimilate best practices in the
medical community Models such as MediTron 44 Apollo
54 and JMLR 51 incorporate guideline data in their
training helping them provide practical advice to healthcare
professionals thus enhancing diagnostic accuracy and reduc
ing the risk of misdiagnosis and mistreatment Despite their
value medical guidelines have limitations They are often
based on average data from the general patient population
which may not always be applicable to individualized clinical
decisionmaking
Package Inserts are documents that provide comprehensive
information about medications and their proper use Intended
for patients doctors and other healthcare professionals they
ensure safe and effective medication use Package inserts
typically include details such as the drug name and ingredients
usage and dosage indications contraindications warnings
and precautions drug interactions and side effects Commonsources for drug insert data include RxList and Drugs These
documents contain rich information that can enhance medical
LLMs understanding of drugs and support decisionmaking
processes However there are limitations Package inserts offer
standardized medication instructions and may not account for
individual patient conditions or concurrent medication use
which can affect the ability of medical LLMs to provide
personalized healthcare and treatment recommendations
Overall the aforementioned public medical corpus sources
offer a wealth of information and extensive coverage pro
viding diverse knowledge resources for medical LLMs Addi
tionally they are relatively easy to access without significant
time and resource investment However these public medical
corpus sources also have some drawbacks Firstly due to the
diversity of medical data sources they need to be filtered
in a finegrained way to obtain data that matches the target
task Secondly most of these data are unstructured and require
appropriate data processing operations to convert them into the
format required for instruction finetuning
3 Professional Medical Organization Corpus Data from
specialized medical institutions such as hospitals and health
care centers is a crucial information source for training
medical LLMs These data includes patients medical histories
diagnostic information treatment plans and doctorpatient
interactions Here we summarize four commonly used pro
fessional medical organization corpus sources Clinical Notes
Electronic Health Records Medical Prescriptions and Doctor
Patient Conversation
Clinical Notes are records created by healthcare profession
als during the consultation diagnosis treatment and moni
toring of patients Typically prepared by doctors nurses and
therapists these notes include patient information symptom
descriptions diagnoses assessments treatment plans and
progress notes MIMIC 91 is a publicly accessible database
of clinical notes offering diverse and comprehensive data
covering the entire course of a patients medical care These
data will enhance medical LLMs abilities in clinical decision
making and case understanding For instance models like
Zhongjing 38 and MeLLaMA 50 use clinical notes for
continued pretraining to enrich their knowledge and under
stand realworld clinical scenarios and reasoning Meanwhile
Clinical Camel 33 aims to enable LLMs to synthesize
plausible clinical records
Electronic Health Records EHRs are digital versions
of patients medical records offering a more comprehensive
overview than clinical notes 92 EHRs include medical
history diagnoses medication regimens immunization dates
allergy information radiology images and laboratory test
results This realtime patientcentric record system provides

sources Specifically medical consultation platforms allow pa

tients to communicate with doctors online for medical advice

diagnoses and treatment plans enhancing the convenience

and accessibility of healthcare services These platforms also

offer doctors a flexible way to provide care more efficiently

Common medical websites include Ding Xiang Yuan 81

Ping An Good Doctor 82 Teladoc Health 83 98point6

84 and Haodf 85 For training medical LLMs LLMs

conversation data from medical consultation platforms is par

ticularly valuable This data is generally more accessible to

the public compared to academic literature and textbooks due

to its realworld colloquial nature Additionally the multi

round questionandanswer format helps improve the models

ability to process contextual information Several models such

as MMedLM 49 ChatDoctor 25 and GPTDoctor 47 use

this type of data to align their response styles with those of real

doctors However data from medical consultation platforms

must be collected with strict adherence to privacy regulations

to ensure patient information security The quality of this

data can vary with some information being inaccurate or

incomplete Therefore it is essential to screen and clean the

data to maintain the accuracy and reliability of the medical

information used for training

Medical Guidelines also known as clinical or practice

guidelines are documents that provide recommendations for

medical decisionmaking including the diagnosis manage

ment and treatment of specific health conditions These

guidelines are critical for medical practice and are typically

developed by professional medical organizations government

agencies or academic groups 86 Common sources include

CCO 87 CDC 88 NICE 89 and WHO 90 Medical

guidelines offer practical information across a broad spec

trum of diseases and clinical scenarios aiding medical LLMs

in learning diagnostic and therapeutic decisionmaking The

standardized format and consensusdriven content of these

guidelines enable models to assimilate best practices in the

medical community Models such as MediTron 44 Apollo

54 and JMLR 51 incorporate guideline data in their

training helping them provide practical advice to healthcare

professionals thus enhancing diagnostic accuracy and reduc

ing the risk of misdiagnosis and mistreatment Despite their

value medical guidelines have limitations They are often

based on average data from the general patient population

which may not always be applicable to individualized clinical

decisionmaking

Package Inserts are documents that provide comprehensive

information about medications and their proper use Intended

for patients doctors and other healthcare professionals they

ensure safe and effective medication use Package inserts

typically include details such as the drug name and ingredients

usage and dosage indications contraindications warnings

and precautions drug interactions and side effects Commonsources for drug insert data include RxList and Drugs These

documents contain rich information that can enhance medical

LLMs understanding of drugs and support decisionmaking

processes However there are limitations Package inserts offer

standardized medication instructions and may not account for

individual patient conditions or concurrent medication use

which can affect the ability of medical LLMs to provide

personalized healthcare and treatment recommendations

Overall the aforementioned public medical corpus sources

offer a wealth of information and extensive coverage pro

viding diverse knowledge resources for medical LLMs Addi

tionally they are relatively easy to access without significant

time and resource investment However these public medical

corpus sources also have some drawbacks Firstly due to the

diversity of medical data sources they need to be filtered

in a finegrained way to obtain data that matches the target

task Secondly most of these data are unstructured and require

appropriate data processing operations to convert them into the

format required for instruction finetuning

3 Professional Medical Organization Corpus Data from

specialized medical institutions such as hospitals and health

care centers is a crucial information source for training

medical LLMs These data includes patients medical histories

diagnostic information treatment plans and doctorpatient

interactions Here we summarize four commonly used pro

fessional medical organization corpus sources Clinical Notes

Electronic Health Records Medical Prescriptions and Doctor

Patient Conversation

Clinical Notes are records created by healthcare profession

als during the consultation diagnosis treatment and moni

toring of patients Typically prepared by doctors nurses and

therapists these notes include patient information symptom

descriptions diagnoses assessments treatment plans and

progress notes MIMIC 91 is a publicly accessible database

of clinical notes offering diverse and comprehensive data

covering the entire course of a patients medical care These

data will enhance medical LLMs abilities in clinical decision

making and case understanding For instance models like

Zhongjing 38 and MeLLaMA 50 use clinical notes for

continued pretraining to enrich their knowledge and under

stand realworld clinical scenarios and reasoning Meanwhile

Clinical Camel 33 aims to enable LLMs to synthesize

plausible clinical records

Electronic Health Records EHRs are digital versions

of patients medical records offering a more comprehensive

overview than clinical notes 92 EHRs include medical

history diagnoses medication regimens immunization dates

allergy information radiology images and laboratory test

results This realtime patientcentric record system provides