tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
authorized users with immediate secure access to patient
information Publicly accessible sources for EHRs include
eICU 93 MIMICIV 94 and CPRD 95 Given the close
relation to medical practice and the multidimensional nature of
the data some works 36 38 have utilized EHR data to train
healthcare domain models enhancing their usefulness and
authenticity However different healthcare organizations may
use varying EHR systems and record formats necessitating
additional preprocessing to standardize the data FurthermoreJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 7
EHRs contain personal medical information that must be
anonymized to protect patient privacy
Medical Prescription is an official directive from a physi
cian or authorized healthcare professional to a pharmacist
detailing the medication to be dispensed to a patient includ
ing the specific drug dosage frequency and administration
method These data reflect realworld prescription practices
providing valuable and authentic information to enhance mod
els decisionmaking in healthcare Some hospitals have inte
grated electronic medical prescriptions into their EHR systems
significantly facilitating scientific research on this data Qibo
55 utilizes Traditional Chinese Medicine TCM prescrip
tions as pretraining corpus to incorporate TCM knowledge into
medical LLMs Medical Prescription data is relatively scarce
and generally requires collaboration with hospitals to access
Additionally these datasets may exhibit biases with certain
hospitals or physicians favoring specific drugs or treatments
Errors omissions or inaccuracies such as misspelled drug
names or incorrect dosages can also compromise the effec
tiveness of model training and application
DoctorPatient Conversation data encompasses reallife
QA and interactions between doctors and patients includ
ing symptom descriptions diagnostic information and pa
tient feedback These conversations typically involve multiple
rounds of communication with doctors actively asking ques
tions Such data can enhance the capabilities of medical LLMs
in assisting diagnosis providing treatment suggestions asking
relevant questions and understanding context HuotuoGPT 8
uses real medical conversation data in serving instruction fine
tuning to enhance long conversations for medical LLMs How
ever accessing this data is challenging due to privacy concerns
from hospitals and patients and the colloquial and emotional
nature of these conversations requires future processing
Overall the aforementioned professional medical organiza
tion corpus sources generated in real medical scenarios pos
sess high authenticity and utility These data provide valuable
information for medical LLMs thereby enhancing their ability
to assist in realworld diagnoses However these data sources
also present several challenges Primarily most of the data
are proprietary to hospitals and require relevant licenses for
access increasing the difficulty and cost of data acquisition
Additionally since the data are generated in offline medical
settings they may be incomplete or inaccurate with potential
errors or gaps in medical records Privacy and security con
cerns are also significant as these data may contain sensitive
patient information raising serious legal and ethical issues if
leaked or misused
4 Synthetic Data Synthetic data is generated by prompt
ing a LLM to create content based on its internal knowledge
or contextual information The common general synthetic
datasets are Alpaca52k ShareGPT which are in QA format
These datasets aim to align ones own LLM with stateof
theart models at minimal cost In the medical domain sev
eral LLMs utilize stateoftheart models to generate training
datasets constructing both singleround and multiround QA
formats
For singleround QA data ChatMedConsult 28 takes
medical questions from the internet as input to ChatGPTand then concatenates its outputs to form an instruction fine
tuning dataset HuatuoGPT 8 employs a selfinstruct 96
method generating a medical instruction dataset by manually
creating seed instructions based on specific roles and use cases
Alpacare 40 initially creates a small set of cliniciancurated
seed tasks considering four dimensions topic view type
and difficulty level Each task includes detailed information
specific to these dimensions instructions and possibly corre
sponding inputs GPT4 then generates new instructions based
on these seed tasks filtering out highly similar ones Finally
ChatGPT answers each generated instruction to compose the
MedInstruct52k instruction finetuning dataset For multi
round QA data HuatuoGPT employs two ChatGPTs one
simulates a doctor and the other as a patient Real data is uti
lized to stimulate conversations between these two ChatGPTs
resulting in a multiround QA dataset
The quality of the generated data depends on the effec
tiveness of prompts Welldesigned prompts can significantly
stimulate the intrinsic knowledge of LLMs resulting in higher
quality data However since this data is not derived from real
world sources manual checking and correction are typically
required to enhance its quality
B Data Processing
After acquiring data from the target source it may not
be immediately suitable for training medical LLMs due to
several issues presence of errors or irrelevant information
nonconforming data formats insufficient or overly brief
content language mismatches between the data source and
target tone inconsistencies with the robotic doctor and lack
of professional annotation Therefore this paper summarizes
the commonly used data processing methods for medical
LLMs and classifies them into four categories according to
their functions namely Data Cleaning Data Formatting Data
Augmentation and Translation

authorized users with immediate secure access to patient

information Publicly accessible sources for EHRs include

eICU 93 MIMICIV 94 and CPRD 95 Given the close

relation to medical practice and the multidimensional nature of

the data some works 36 38 have utilized EHR data to train

healthcare domain models enhancing their usefulness and

authenticity However different healthcare organizations may

use varying EHR systems and record formats necessitating

additional preprocessing to standardize the data FurthermoreJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 7

EHRs contain personal medical information that must be

anonymized to protect patient privacy

Medical Prescription is an official directive from a physi

cian or authorized healthcare professional to a pharmacist

detailing the medication to be dispensed to a patient includ

ing the specific drug dosage frequency and administration

method These data reflect realworld prescription practices

providing valuable and authentic information to enhance mod

els decisionmaking in healthcare Some hospitals have inte

grated electronic medical prescriptions into their EHR systems

significantly facilitating scientific research on this data Qibo

55 utilizes Traditional Chinese Medicine TCM prescrip

tions as pretraining corpus to incorporate TCM knowledge into

medical LLMs Medical Prescription data is relatively scarce

and generally requires collaboration with hospitals to access

Additionally these datasets may exhibit biases with certain

hospitals or physicians favoring specific drugs or treatments

Errors omissions or inaccuracies such as misspelled drug

names or incorrect dosages can also compromise the effec

tiveness of model training and application

DoctorPatient Conversation data encompasses reallife

QA and interactions between doctors and patients includ

ing symptom descriptions diagnostic information and pa

tient feedback These conversations typically involve multiple

rounds of communication with doctors actively asking ques

tions Such data can enhance the capabilities of medical LLMs

in assisting diagnosis providing treatment suggestions asking

relevant questions and understanding context HuotuoGPT 8

uses real medical conversation data in serving instruction fine

tuning to enhance long conversations for medical LLMs How

ever accessing this data is challenging due to privacy concerns

from hospitals and patients and the colloquial and emotional

nature of these conversations requires future processing

Overall the aforementioned professional medical organiza

tion corpus sources generated in real medical scenarios pos

sess high authenticity and utility These data provide valuable

information for medical LLMs thereby enhancing their ability

to assist in realworld diagnoses However these data sources

also present several challenges Primarily most of the data

are proprietary to hospitals and require relevant licenses for

access increasing the difficulty and cost of data acquisition

Additionally since the data are generated in offline medical

settings they may be incomplete or inaccurate with potential

errors or gaps in medical records Privacy and security con

cerns are also significant as these data may contain sensitive

patient information raising serious legal and ethical issues if

leaked or misused

4 Synthetic Data Synthetic data is generated by prompt

ing a LLM to create content based on its internal knowledge

or contextual information The common general synthetic

datasets are Alpaca52k ShareGPT which are in QA format

These datasets aim to align ones own LLM with stateof

theart models at minimal cost In the medical domain sev

eral LLMs utilize stateoftheart models to generate training

datasets constructing both singleround and multiround QA

formats

For singleround QA data ChatMedConsult 28 takes

medical questions from the internet as input to ChatGPTand then concatenates its outputs to form an instruction fine

tuning dataset HuatuoGPT 8 employs a selfinstruct 96

method generating a medical instruction dataset by manually

creating seed instructions based on specific roles and use cases

Alpacare 40 initially creates a small set of cliniciancurated

seed tasks considering four dimensions topic view type

and difficulty level Each task includes detailed information

specific to these dimensions instructions and possibly corre

sponding inputs GPT4 then generates new instructions based

on these seed tasks filtering out highly similar ones Finally

ChatGPT answers each generated instruction to compose the

MedInstruct52k instruction finetuning dataset For multi

round QA data HuatuoGPT employs two ChatGPTs one

simulates a doctor and the other as a patient Real data is uti

lized to stimulate conversations between these two ChatGPTs

resulting in a multiround QA dataset

The quality of the generated data depends on the effec

tiveness of prompts Welldesigned prompts can significantly

stimulate the intrinsic knowledge of LLMs resulting in higher

quality data However since this data is not derived from real

world sources manual checking and correction are typically

required to enhance its quality

B Data Processing

After acquiring data from the target source it may not

be immediately suitable for training medical LLMs due to

several issues presence of errors or irrelevant information

nonconforming data formats insufficient or overly brief

content language mismatches between the data source and

target tone inconsistencies with the robotic doctor and lack

of professional annotation Therefore this paper summarizes

the commonly used data processing methods for medical

LLMs and classifies them into four categories according to

their functions namely Data Cleaning Data Formatting Data

Augmentation and Translation