Skip to content

Latest commit

 

History

History
199 lines (155 loc) · 4.79 KB

README.md

File metadata and controls

199 lines (155 loc) · 4.79 KB
Raw

ppg2nrs

Pre&Post-op ppg > NRS (post-op, in PACU)

Usage

dataset class

from dataset import VitalDataset_fs
root_dir = f'../data/all_3'
trdt  = VitalDataset_fs(root_dir,f'../data/pd_gy/train_3.json')
valdt = VitalDataset_fs(root_dir,f'../data/pd_gy/val_3.json') 
tedt  = VitalDataset_fs(root_dir,f'../data/pd_gy/test_3.json')

model class

from models import PrePostNet
model = PrePostNet()

train and validation

from train import trainval
best_model,val_losses = trainval(trdl,valdl,model,loss,opt,scheduler=scheduler,device=device, exist_acc=True)

Input

Pre-op(5Min after NIBP(non-invasive blood pressure))

  • ECG
  • PPG Post-op(5Min)
  • ECG
  • PPG

Output

label(pain) nrs
0 0~3
1 4~9

Preprocessing

sample_rate: 300Hz

PPG

  • interpolate (constant)
  • Bandpass filter (0.5~15)
  • Moving average (30taps) (- z-score)
  • Resampling (to set 90000)
  • Spectrogram

ECG

  • interpolation (constant)
  • elgendi preprocessing

Spectogram for ECG and PPG (of pre/post-op)

  • ppg
  • ecg

Model

model input structure

Type 1. 4 channel

Type 2. OR vs REC

Data

  • label
    0: no-milf psin, 1: moderate-severe pain
nrs2
0 545
1 116

Fast-dataset vs Original

There are two kinds of dataset in dataset.py
The fast-dataset extracted final results as a numpy from original dataset.

Data Version History

v1

======== x: big vital files

  • vital files paths 모두 추출 (#1) ======== y: label file (회복간호_21상, 회복간호_21하, 회복간호_22상) -통합(df),

  • nrs없는 항목 제거,

  • (등록번호, 회복실퇴실날짜) 중복 시 제거

  • '등록번호','회복실퇴실일시','최대 NRS','KEY' 항목들만 추출

  • 회복실퇴실일시> time, date 추출 (time, date를 key처럼 사용 (vital file 추출 용이))

    • df: ['등록번호','date','time','최대 NRS','KEY']
    • df: ['pt_id', 'date', 'time', 'max_nrs','key']

  • #1를 이용하여 df의 해당하는 vital이 없을 경우 삭제

    • df: 'idx', 'label', 'vital', 'path', 'nrs', 'key', 'ext_path', 'room', 'path_0'

  • 섞여있는 필요한 데이터를 path를 이용하여 알맞게 복사 (#data moving)

  • 해당 수술에 대한 pdor과 pdrec을 매칭시켜 저장 (df > json) #json scheme ======== #data moving

  • 사용되는 흩어진 데이터 > pdor/*, pdrec/에 복사 (18GB, 약 1600pair)
    // 파일의 이름을 이용하여 pdor, pdrec 구별 ======== #json scheme
    item[N]: 'key', 'rec_path', 'or_path', 'nrs' key: 해당 key는 raw data(회복간호    .csv) 기재된 키. 이를 통해 환자의 다른 메타정보를 추적할 수 있음.
    rec_path: 회복실 vital file paths
    or_path: 수술시 vital file paths
    nrs: (numeric rating score) 정답 라벨 // 통증의 정도를 0~10범위로 표현, 회복실에서 측정됨.

======== Final Scheme
./pdrec/* (# data moving)
./pdor/*
./all.json (# json scheme)

v1.5

  • +demographic data (gender, age)

======== Final Scheme

./pdrec/* (# data moving)
./pdor/* ./all_1.json

v3

  • Applied Filter4
    1. NRS?
    2. pdor&pdrec pair?
    3. General Anesthesia?
    4. Gynecology dept.?

======== Final Scheme ./pdrec/* (# data moving)
./pdor/* ./all_2.json

======== Data DESC.
About label
no-mild: moderate-severe = 545:116 ≓ 4.5:1

v4

  • Applied Filter4
    1. NRS?
    2. pdor&pdrec pair?
    3. General Anesthesia?
    4. Gynecology dept.?
    5. Data Quality? (Non-existing, too many null, …) (black list)

======== Final Scheme
./all_3.json
../data/all_3/*.npy

./train_3.json
./val_3.json
./test_3.json

black list

black_list = ['01826958_PDOR1_210421_124500_1',
              '01853529_PDOR1_210419_083500_1',
              '01876359_PDOR2_210602_083200_1',
              '01656320_PDOR2_210526_091300_1',
              '01618678_PDOR1_210615_083000_1',
              '01159584_PDOR1_211025_083400_1',
              '00578148_PDOR2_210805_111500_1',
              '01853408_PDOR1_211115_083100_1',
              '01906040_PDOR1_210831_110500_1',
              '00435676_PDREC02_210902_093500_1', # dataset 
              '01060037_PDOR2_211206_102000_1', #  Filling missing values error
              '01929835_PDOR2_211102_093000_1',
              '01392754_PDOR2_220111_105500_1',
              '01439205_PDOR2_211206_121100_1', # in dl, fmissing
              '00435676_PDOR1_210902_084000_1'
             ]

Acknowledge

Dongheon Lee , Ph.D Boohwi Hong , M.D., Ph.D.