Progress Seminar 2018.01.25 이준녕
연구 진행 상황 보고서 재활의학과 응급의학과 혈압팀 2주전 계획 연구 결과 문제점 및 대책 목표 및 계획 - Physionet data RNN 분석 연구 결과 Physionet data RNN 분석 (dropout 추가, peep-hole connection 확인, 다양한 architecture 성능 확인) >500,000 데이터 분석 논문 틀 작성 문제점 및 대책 Class balancing + subject balancing 목표 및 계획 특허 작성 논문 작성
MIMIC data pre-processing
MIMIC data BP estimation features
RNN input-output structure
Confusion matrix
17 subjects, 86 recordings, >24610 samples 60<DBP<90 (24610 samples) 100<SBP<140 (15420 samples) Comparison with BHS standard Comparison with AAMI standard (may be meaningless) <5mmHg <10mmHg <15mmHg SBP 57.3 87.3 94.0 DBP 67.1 92.6 96.7 Grade A 60% 85% 95% ME STD Subjects SBP -0.93 9.72 86 (or 17?) DBP -0.94 8.41 SBP/DBP <5 <8 >85
168 subjects, 945 recordings, >140000 samples 60<DBP<90 (24610 samples) 100<SBP<140 (15420 samples) Comparison with BHS standard Comparison with AAMI standard (may be meaningless) <5mmHg <10mmHg <15mmHg SBP DBP 47.8 73.7 84.1 Grade A 60% 85% 95% Grade B 50 75 90 Grade C 40 65 85 ME STD Subjects SBP DBP 1.41 13.5 168 SBP/DBP <5mmHg <8mmHg n>85
Validation 문제점 Current method: Balance #of data per class: results in different # data per subject 80% of randomly chosen data for training, 20% for testing Data from 1 subject may be in both training and testing Solution 1: Balance # of data per class & # data per subject -> not enough data for training -> may be possible to use only subjects with large # of data Solution 2: pick random 20% of subjects from ~165 subject pool -> unbalanced classes in training and testing data Solution 3: balance # of data per subject -> unbalanced # of data per class Solution 4: add balanced # of subjects & data for validation p01 of the dataset