Risk score-embedded deep learning for biological age estimation: Development and validation


Suhyeon Kim, Hangyeol Kim, Eun-Sol Lee, Chiehyeon Lim, 이정혜 (2022) · information-sciences 586:628-643 · DOI ↗

RSAE-BA (Risk Score-Embedded Autoencoder for Biological Age)health risk 정보를 오토인코더 embedding 의 loss function 에 명시 통합biological age (BA) 추정 새 알고리즘. 기존 BA 추정의 3 한계 극복: (1) Linear regression (MLR, lasso, elastic net) — BA 를 CA (chronological age) 로 distort, (2) Deep learning regression (Cole et al 2017 CNN, Pyrkov et al 2018) — CA 를 target 으로 supervised — BA goal 과 deviation, (3) Klemera-Doubal Method (KDM 2006) + PCA — assumption-heavy or variance-maximization directionhealth-irrelevance. 4 단계: (i) Cumulative Gaussian probability 기반 Risk Score (RS) 계산 — Type 1 (large=poor), Type 2 (small=poor), Type 3 (both=poor); (ii) Data transformation 으로 방향 정렬; (iii) RSAE modeling — loss LRSAE=LKL(X,X)+LKL(X,P)\mathcal{L}_{RSAE} = \mathcal{L}_{KL}(\mathbf{X}', \mathbf{X}'') + \mathcal{L}_{KL}(\mathbf{X}'', \mathbf{P}); (iv) BA = TT-scale 변환된 standardized BAS. 새 BA validation — LR vs HR group 의 RS 차이로 unlabeled + labeled 모두 평가. 3 dataset 검증: KNHANES (85,490명, unlabeled), NHIC-HTN (140,867명, 2.5 년 hypertension 발생), NHIC-T2DM (T2DM 발생). 이정혜3 기 TEMEP 시기 의료 AI 정점.

  • RQ: (1) BA 추정에 health risk 정보를 deep representation learning embedding 에 통합할 수 있는가? (2) BA validation 의 cohort dependency 한계 (mortality/disease incidence 필요) 를 어떻게 unlabeled data 에도 적용 가능하게 만드는가? (3) RSAE-BA 가 기존 방법 (MLR, lasso, KDM, PCA, vanilla AE) 대비 질병 incidence 예측 에 우위인가?
  • 방법론: Stage 1 — Risk Score (RS) calculation: 각 변수의 normal 분포 N(μj,σj2)\mathcal{N}(\mu_j, \sigma_j^2) (μj\mu_j = medical prior, σj\sigma_j = empirical) 의 cumulative probability. 3 type: Type 1 (one-sided large), Type 2 (one-sided small), Type 3 (two-sided). Stage 2 — Data transformation (Algorithm 1): standardize → Type-specific direction alignment (×-1 or |·|) → sigmoid 로 [0,1] bounding. Stage 3 — RSAE modeling (Algorithm 2): encoder HEnc(k)=fReLU(HEnc(k1)WEnc(k)+bEnc(k))\mathbf{H}_{Enc}^{(k)} = f_{ReLU}(\mathbf{H}_{Enc}^{(k-1)} \mathbf{W}_{Enc}^{(k)} + \mathbf{b}_{Enc}^{(k)}), decoder mirror. KL divergence loss = reconstruction + RS-prediction. Stage 4 — BA calculation: BAS = zz 의 평균 standardize, BA = TT-scale transformation. BA validation — Group 1 (BA<CA) 의 LR group (q%q\% largest CA−BA) vs Group 2 (BA>CA) 의 HR group (q%q\% largest BA−CA); hypothesis: RSLR<RSHRRS_{LR} < RS_{HR}, good BA = larger RS gap
  • 데이터: (i) KNHANES (Korea National Health and Nutrition Examination Survey, 1988-2017): 85,490 명 — unlabeled. (ii) NHIC-HTN (Korean National Health Insurance Corporation 2002-2010, hypertension cohort): 140,867 명 — 2.5 yr HTN 발생 labeled. (iii) NHIC-T2DM (type-2 diabetes cohort): T2DM 발생 labeled. 2% population sample stratified by age, sex, income. Variables: clinical biomarker (BP, glucose, cholesterol, BMI 등)
  • 주요 발견: (i) RSAE-BA > MLR / lasso / KDM / PCA-BA / vanilla AE-BA 에서 모든 데이터셋 (KNHANES, NHIC-HTN, NHIC-T2DM) 의 LR vs HR group RS gap 최대. (ii) Disease incidence 예측 — RSAE-BA 가 HTN, T2DM 발생 의 AUC/AUROC baseline 대비 향상. (iii) Unlabeled data 검증 가능 — RS 기반 validation 이 cohort dependency 한계 해소. (iv) 3 Type RS classification 의 정합성 — 모든 clinical variable 의 risk direction 명시. (v) Sample size 85k → 140k 의 scalability 검증
  • 시사점: (a) Hyper-personalized health managementquantitative anchor — BA 를 useful health index 로 일상 활용. (b) RS embedding pattern 의 generalizability — 다른 health index (e.g. mental health score), 다른 domain (e.g. household financial risk of suhyeon-kim-2023-household-financial-rihae) 로 확장. (c) Cohort-independent validation — 일반 EMR (electronic medical record) 만으로 BA index 평가 가능. (d) Medical prior + data-driven hybrid — Gaussian μj\mu_jmedical normal criterion + σj\sigma_jempirical 결합. (e) 예방 의학 + 만성질환 (HTN, T2DM) 의 early warning index

Fig. 3 의 RSAE 구조: input \mathbf{X}' → encoder (K layers, ReLU) → bottleneck \mathbf{Z} → decoder → reconstruction \mathbf{X}''. Loss \mathcal{L}_{RSAE} = \mathcal{L}_{KL}(\mathbf{X}', \mathbf{X}'') + \mathcal{L}_{KL}(\mathbf{X}'', \mathbf{P}) 의 dual term 이 \mathbf{Z} 에 reconstruction information + health risk information 동시 인코딩. RS \mathbf{P} 가 latent space 에 explicit embedding.

요약

본 paper 는 이정혜3 기 SNU TEMEP 부임 초기 (2022 published, 2021 work)의료 AI 정점. Suhyeon Kim (W2V-LSA, LBC 의 동일 제 1 저자) + Hangyeol Kim + Eun-Sol Lee + UNIST Chiehyeon Lim (Department of Industrial Engineering & Graduate School of Artificial Intelligence, Bio-Age Inc 의 medical institute affiliation) + 이정혜 (corresponding) 의 UNIST-Bio-Age industry-academia 협업. 동기는 Chronological Age (CA) 의 한계 — 시간 흐름의 단순 측정으로 functional, structural body capacity + aging 평가 부정확. Biological Age (BA) 가 대안 health index — EMR 요약, 연령-의존 biomarker 결합. 기존 3 type BA 의 systematic 한계:

  1. Regression for CA-based (MLR, lasso, elastic net + CNN deep learning) — BA-to-CA distortion + CA 가 target 인 supervised setting 의 deviation
  2. Simulation-based (KDM, Klemera-Doubal 2006) — multidimensional space 의 regression line - biomarker point 거리 minimize — assumption-heavy, mortality 예측 우월하지만 data conformity 의존
  3. PCA-based latent feature — variance-maximize 방향이 health-relevance 와 무관

또한 모든 기존 방법이 risk direction (variable 의 bad level 이 large vs small 인지) 무시. Validationcohort data 의 mortality/incidence label 의존 — 일반 EMR 적용 어려움. 본 paper 는 RS-embedded representation learning 로 두 한계 동시 해결.

Methodology Stage 1 — Risk Score (RS) calculation. 각 변수 xj\mathbf{x}_j 의 normal 분포 N(μj,σj2)\mathcal{N}(\mu_j, \sigma_j^2)μj\mu_j = medical examination normal criterion (prior knowledge), σj2\sigma_j^2 = data empirical. RS = cumulative Gaussian probability. 3 type:

  • Type 1: large = poor (예: glucose, BP). One-sided right.
  • Type 2: small = poor (예: HDL cholesterol). One-sided left.
  • Type 3: large or small = poor (예: heart rate). Two-sided.

다른 분포 (binomial, multinomial) 로 extensible.

Stage 2 — Data transformation (Algorithm 1). standardize → Type-specific direction alignment (Type 2: ×1\times -1; Type 3: |\cdot|) → sigmoid s(a)=1/(1+ea)s(a) = 1/(1+e^{-a})[0,1][0,1] bounding. Direction 정렬이 loss accumulation error 방지.

Stage 3 — RSAE modeling (Algorithm 2). KK hidden layer encoder + mirror decoder, ReLU activation. Loss:

LRSAE=LKL(X,X)+LKL(X,P)\mathcal{L}_{RSAE} = \mathcal{L}_{KL}(\mathbf{X}', \mathbf{X}'') + \mathcal{L}_{KL}(\mathbf{X}'', \mathbf{P})

여기서 X\mathbf{X}' = transformed input, X\mathbf{X}'' = reconstruction, P\mathbf{P} = RS matrix. KL divergence 가 X\mathbf{X}'' 를 reconstruction + RS 양쪽에 fit. Multi-task learning 이 bottleneck Z\mathbf{Z}generic latent + health risk 동시 인코딩.

Stage 4 — BA calculation. BA Score (BAS) = zdz_d 의 평균 standardize:

BAS=1D(K)d=1D(K)(zdzdμ1n)(zdσ1n)\text{BAS} = \frac{1}{D^{(K)}} \sum_{d=1}^{D^{(K)}} (\mathbf{z}_d - z_d^\mu \mathbf{1}_n) \oslash (z_d^\sigma \mathbf{1}_n)

BA = TT-scale transformation (standard score → TT-score) 으로 year scale 의 BA.

BA validation (novel). 개인 sort: Group 1 (BA<CA), Group 2 (BA>CA). LR (low-risk) group = Group 1 의 q%q\% largest CA−BA. HR (high-risk) group = Group 2 의 q%q\% largest BA−CA. 가설: (i) RSLR<RSHRRS_{LR} < RS_{HR}, (ii) good BA = larger RS gap. Unlabeled + labeled 모두 적용 가능 — cohort 의존성 해소.

Experiments — 3 dataset:

  • KNHANES (85,490명, unlabeled): RS-based validation
  • NHIC-HTN (140,867명, hypertension cohort): 2.5 yr HTN 발생 prediction
  • NHIC-T2DM: type-2 diabetes 발생 prediction

결과: RSAE-BA 가 모든 dataset 에서 baseline (MLR-BA, lasso-BA, KDM, PCA-BA, vanilla AE-BA) 대비 LR-HR RS gap 최대 + disease incidence 예측 AUROC 최대. 한계: (i) Gaussian normal 가정 의 all variable applicability, (ii) q%q\% 의 ad-hoc selection, (iii) 한국 cohort 한정 — 다른 인구 generalizability 미검증, (iv) hyperparameter (KK, D(K)D^{(K)}, learning rate) tuning data-dependent. 이정혜 author page 의 3 기 의료 AI 라인의 peak + suhyeon-kim 4-paper seriesmedical AI representative.

핵심 결과

3 dataset 검증

DatasetSampleType사용
KNHANES85,490Unlabeled (annual survey 1988-2017)RS-based BA validation
NHIC-HTN140,867Labeled (2.5 yr HTN 발생)Hypertension prediction
NHIC-T2DM(similar)Labeled (T2DM 발생)Type-2 diabetes prediction

RS 3 type classification

TypeDirectionDistribution
1Large = poorOne-sided (right)Glucose, Blood Pressure
2Small = poorOne-sided (left)HDL Cholesterol
3Both = poorTwo-sidedHeart rate, BMI

비교 baseline (Section 3)

Method약점
MLR-BACA distortion
Lasso-BA, Elastic Net-BA동일
KDM (Klemera-Doubal)Assumption-heavy
PCA-BAVariance direction ≠ health
Vanilla AE-BAGeneric latent, no health risk
RSAE-BA모든 결함 해결

정량 결론. RSAE-BA 가 모든 데이터셋에서 LR-HR RS gap 최대 + disease incidence AUROC 최대. RS-embedded loss 의 dual-term 이 latent space 에 health risk 명시 인코딩. Cohort-independent validation 가능.

방법론 노트

Risk Score: 변수 xj\mathbf{x}_j 의 RS pj[0,1]n\mathbf{p}_j \in [0,1]^n = Gaussian N(μj,σj2)\mathcal{N}(\mu_j, \sigma_j^2) 의 cumulative probability. Type 별 one-sided 또는 two-sided.

RSAE encoder (Eq 2):

HEnc(k)=fReLU(HEnc(k1)WEnc(k)+1n(bEnc(k))T)\mathbf{H}_{Enc}^{(k)} = f_{ReLU}(\mathbf{H}_{Enc}^{(k-1)} \mathbf{W}_{Enc}^{(k)} + \mathbf{1}_n \cdot (\mathbf{b}_{Enc}^{(k)})^T)

RSAE loss (Eq 6) — KL divergence dual term:

LRSAE=LKL(X,X)+LKL(X,P)\mathcal{L}_{RSAE} = \mathcal{L}_{KL}(\mathbf{X}', \mathbf{X}'') + \mathcal{L}_{KL}(\mathbf{X}'', \mathbf{P})

여기서 LKL(x,y)=ixilog(xi/yi)\mathcal{L}_{KL}(x, y) = \sum_i x_i \log(x_i / y_i).

BA calculation (Eq 7):

BAS=1D(K)d=1D(K)(zdzdμ1n)(zdσ1n)\text{BAS} = \frac{1}{D^{(K)}} \sum_{d=1}^{D^{(K)}} (\mathbf{z}_d - z_d^{\mu} \mathbf{1}_n) \oslash (z_d^{\sigma} \mathbf{1}_n)

TT-scale: Ti=10BASi+50T_i = 10 \cdot \text{BAS}_i + 50 (또는 CA 와 align). BA in years.

BA validation:

LR={iGroup 1CAiBAitop q%}\text{LR} = \{i \in \text{Group 1} \mid \text{CA}_i - \text{BA}_i \in \text{top } q\%\} HR={iGroup 2BAiCAitop q%}\text{HR} = \{i \in \text{Group 2} \mid \text{BA}_i - \text{CA}_i \in \text{top } q\%\}

Test: RSˉLR<RSˉHR\bar{RS}_{LR} < \bar{RS}_{HR} + gap 의 크기.

식별은 (i) 85k-140k sample 의 large-scale variation, (ii) 3 dataset 의 unlabeled + labeled 동시 검증, (iii) baseline 다양성 (linear, simulation, latent feature, vanilla AE), (iv) RS embedding 의 interpretable 의미 (Gaussian + medical prior). Limitation: Gaussian 가정 + ad-hoc qq + 한국 cohort 한정.

연구 계보

본 paper 의 BA estimation lineage: Hannum et al (2013) DNA methylation BA, Levine et al (2018) PhenoAge, Klemera-Doubal (2006 Mechanisms of Aging and Development, KDM), Cole et al (2017 NeuroImage, CNN brain age), Pyrkov et al (2018 Scientific Reports), Rahman-Adjeroh (2019). Health index lineage: Mitnitski-Mogilner-Rockwood (2001 frailty index), DeAngelis et al (2009). Autoencoder: Hinton-Salakhutdinov (2006 Science), Bengio et al (2013 PAMI), Vincent et al (2010 denoising AE). Risk Score 의 cumulative Gaussian: medical statistics 전통 — Z-score, percentile rank. KL divergence: Kullback-Leibler (1951).

TEMEP 내 sibling: Suhyeon Kim4-paper seriesmedical AI peakWord2vec-based latent semantic analysis (W2V-LSA) for topic modeling: A study on blockchain technology trend analysis (text embedding) → A Multi-stage Data Mining Approach for Liquid Bulk Cargo Volume Analysis based on Bill of Lading Data (logistics) → 본 paper (medical AI) → TMF-GNN: Temporal matrix factorization-based graph neural network for multivariate time series forecasting with missing values (graph neural network). 직접 후속: RI-HAE (Risk-Information Hybrid Autoencoder) 의 household financial risk domain extension — RSAE 의 architecture 가 medical + financial 도메인 모두 transferable. 이정혜 author page 의 실타래 3 (Healthcare AI) + 실타래 4 (Representation Learning) 합류 의 anchor. Chiehyeon Lim (UNIST AI graduate school) 과의 협업 — UNIST 의 Bio-Age Inc 산학 cluster 의 academic output.

See also

인접 그래프

1-hop 이웃 15
  • 인물 5
  • 방법론 3
  • 주제 1
  • 논문 6
휠 = 확대/축소 · 드래그 = 이동 · hover = 강조 · 클릭 = 페이지 이동