📑 챕터 바로가기
PART 2는 "추정(estimation·Nowcasting)"이다. 새 사람이 와도 — 설문과 외부 코호트(KNHANES)로 평균 수준 L을 잡고, 하루의 위상(일주기)을 더해 구조적 추정치 E0 = 수준 + 위상을 만든다. 그 뒤 남는 편차 r = 실측 − E0를 워치(급성 생리)로 줄인다. 과거 BP추정변수는 쓰지 않고, days2-7 전부를 추정에 활용한다(예측구간 미분리). 최종은 분단위 혈압 곡선이며, 성과는 실측 시점에서만 잰다(MAE + BHS, 세션·일평균).
구조
추정의 계층 구조 — 수준 · 위상 · 편차
- 수준 L(p). 외부(KNHANES 가중회귀)+설문+day1 보정을 shrinkage로 합친 사람별 평균 혈압 — 2-1·2-2·2-3
- 위상. 일주기 phase 커널. 수준과 합쳐 E0 = 수준 + 위상 (구조적 추정치) — 2-4
- 편차 r = 실측 − E0. 워치(급성 생리)로 남은 편차를 모델(과거 BP추정변수 제외) — 2-5 · 편차모델
- 최종 = E0 + r̂. 분단위로 출력, 세션·일평균으로 집계, 실측에서 평가
E0(t) = L(p) + 위상(tod) | r(t) = 실측(t) − E0(t) ≈ f(워치(t)) | 최종(t) = E0(t) + r̂(t)
평가 원칙. 학습·집계엔 보간값을 모두 쓰되, 성과는 실측 시점에서만 MAE + BHS로. 예측구간을 따로 두지 않고 days2-7 전부를 추정에 활용한다(day1은 E0 보정 앵커).
2-1
외부데이터 — KNHANES 앵커
신규 환자는 과거 혈압이 없다. 그래서 국민건강영양조사(KNHANES)를 외부 앵커로 — 나이·성별·BMI·설문 변수로 가중회귀해 그 사람의 모집단 기대 수준 Lext를 만든다. 내부 코호트가 작아도(120명) 외부 분포로 수준을 안정화하는 shrinkage의 바깥쪽 닻이다.
수준 L = (n·cm + 2·Lext) / (n + 2) — day1 보정평균 cm을 Lext로 shrink (n=day1 측정수)
2-2
내부데이터 — 장비 offset 정렬
KNHANES는 표준 cuff, 우리는 다른 장비·프로토콜이다. 둘의 측정 scale 차이를 내부 train 코호트로 offset α를 적합해 정렬한다 — Lext를 우리 장비 단위로 옮기는 보정. 내부데이터는 또한 위상 커널·편차 모델의 학습원이다(train 사람 days2-7).
2-3
1-day calibration — 개인 닻
day1(첫날) 측정만 보정에 쓴다(이후 실측 과거 BP는 금지). day1 평균 cm을 외부 수준 Lext로 shrink해 사람별 level을 잡는다. day1이 많을수록(n↑) 개인값을, 적을수록 외부 앵커를 더 믿는 경험적 베이즈 구조.
수준 분해(검증 24명, days2-7): 모집단 평균에서 출발해 외부+내부offset → +1day → +위상으로 갈수록 일평균 MAE가 단계적으로 낮아진다(2-4 표).
2-4
위상 커널 — 일주기 추정
혈압은 하루 주기로 오르내린다(낮 높고 새벽 낮음). 같은 시간대(time-of-day)일수록 큰 가중치를 주는 phase 커널로 일주기를 복원해 수준에 더한다. 하루는 1440분 원이라 circular 거리를 쓴다.
위상(tod) = Σ exp(−circ(φ,φ')²/2ℓ²)·dev(φ') / Σ exp(...) , circ(a,b)=min(|a−b|,1440−|a−b|)
❗ 여기가 핵심 — 보간과 추정에서 위상은 "무엇에" 기대는지가 다르다.
· 보간(PART1): 그날 자기 실측점을 좁은 커널(ℓ=90·κ=0.5)로 연결 → 실제 그날 리듬 재현.
· 추정(PART2): 신규 환자는 그날 측정이 하나도 없다. 그래서 위상을 E0 = 수준 + [α·(day1 위상·최근성 가중 편차) + (1−α)·(집단 평균 일주기)] 로 만든다. day1 점을 recency(τ)×phase(ℓ=196)로 가중하되, α = W/(W+κ), κ=5라 확신이 낮고, 목표가 day1에서 멀수록 recency가 죽어 α→0 → 사실상 집단(연령×성별) 평균 일주기를 차용한다.
· 보간(PART1): 그날 자기 실측점을 좁은 커널(ℓ=90·κ=0.5)로 연결 → 실제 그날 리듬 재현.
· 추정(PART2): 신규 환자는 그날 측정이 하나도 없다. 그래서 위상을 E0 = 수준 + [α·(day1 위상·최근성 가중 편차) + (1−α)·(집단 평균 일주기)] 로 만든다. day1 점을 recency(τ)×phase(ℓ=196)로 가중하되, α = W/(W+κ), κ=5라 확신이 낮고, 목표가 day1에서 멀수록 recency가 죽어 α→0 → 사실상 집단(연령×성별) 평균 일주기를 차용한다.
왜 이렇게 / 무슨 뜻인가. 그날 측정이 없으니 희소한 day1 + 안정적 집단 평균에 의존한다(넓은 ℓ·약신뢰 κ). 일주기 형태가 군집에 거의 불변임을 확인했기에 집단 평균이 타당한 prior다. 다만 이건 평균 리듬이라 그날 특유의 급변은 못 담는다 → 세션(순간) 추정의 한계로 이어진다. E0 = 수준 + 위상만으로도 days2-7 test 일평균이 BHS A(일 SBP 2.41 A·DBP 2.35 A); 이후 워치 편차 모델이 세션 DBP·일평균을 더 깎는다.
2-5
와치 — 변수 하나하나 (핵심)
편차의 급성 동인은 워치다. 엑셀 26시트(Info·BP + 24개 신호)를 한 장씩 점검했다. 신호마다 측정 빈도·해상도가 천차만별 — 분단위급(HR·에너지·METs·걸음)부터 sparse(HRV·SpO2·손목온도)까지. 아래는 120명 집계 기초통계.
| 워치 변수 | 단위 | 측정/주·명 | 커버% | 평균±SD | 범위 | 밀도 |
|---|---|---|---|---|---|---|
| 심박수 HR | BPM | 4,462 | 100 | 75.2±13.5 | 43–141 | 고밀도 |
| 활동 에너지 | kcal | 3,753 | 100 | 3.9±3.5 | 0–32 | 고밀도 |
| 휴식 에너지 | kcal | 3,720 | 100 | 17.4±8.2 | 0.1–86.6 | 고밀도 |
| 절대적 운동강도 METs | MET | 3,361 | 100 | 1.8±1.0 | 1–12 | 고밀도 |
| 걷기+달리기 거리 | km | 2,242 | 100 | 0.1±0.1 | 0–0.7 | 고밀도 |
| 걸음 | steps | 1,175 | 100 | 96.6±124.2 | 1–1056 | 고밀도 |
| 일어서기 시간(분) | 분 | 357 | 100 | 1.7±1.0 | 1–5 | 중밀도 |
| 호흡수 | /min | 273 | 99 | 15.5±2.8 | 7.5–32.5 | 중밀도 |
| 운동하기 시간 | 분 | 239 | 97 | 1.0±0.0 | 1 | 중밀도 |
| 보행 보폭 / 속도 | cm·m/s | 213 | 98 | 63.4 / 1.1 | 25–132 | 중밀도 |
| 수면 | 단계 | 205 | 99 | 3.2±1.2 | 0–5 | 중밀도 |
| 이중 지지 시간 | s | 190 | 98 | 0.3±0.0 | 0.2–0.4 | 중밀도 |
| 보행 비대칭성 | % | 102 | 98 | 0.04±0.1 | 0–1 | 중밀도 |
| 혈중 산소 SpO2 | % | 113 | 100 | 1.0(=100%) | 0.8–1.0 | 중밀도 |
| 심박 변이 HRV | ms | 80 | 100 | 45.2±24.8 | 0–221 | 저밀도 |
| 오른 층수 | 층 | 34 | 94 | 1.9±1.5 | 1–17 | 저밀도 |
| 계단 올라가기 속도 | m/s | 21 | 87 | 0.3±0.1 | 0.2–1.1 | 저밀도 |
| 일광 시간 | 시간 | 20 | 78 | 2.9±1.5 | 1–5 | 저밀도 |
| 휴식기 / 걷기 심박수 | BPM | 7 | 100 | 63.1 / 98.5 | 44–146 | 저밀도 |
| 손목 온도 | °C | 3 | 50 | 35.6±0.5 | 34.3–38.2 | 저밀도 |
| 보행 안정성 | - | 0 | 76 | 0.9±0.1 | 0.7–1.0 | 저밀도 |
점검 포인트. 손목온도는 커버리지 50%·주 3회로 가장 sparse하나 말초저항(TPR)의 대리로 이론적 중요도가 높다. SpO2는 decimal(1.0=100%) 코딩. 걸음은 평균보다 SD가 커(우편향) log 후보. 고밀도 신호는 분단위 feature로, sparse는 imputation으로 다룬다.
생성과정 — 원신호를 모형 변수로 (어떻게 만드나)
편차 모델에 실제로 들어가는 8개 워치 변수는 원신호를 이렇게 가공해 만든다(쉬운 말로).
- 심박수 HR — 그 세션의 마지막 심박수 값. 가장 기본 신호.
- 가중 심박수 kwHR — 그 시점 주변 심박수를 커널로 평균낸 값(순간 튐을 눌러 매끈하게).
- ΔHR(안정 대비) dHR — HR − 안정심박. 안정심박은 그 사람 HR의 하위 10%(휴식 수준)로 잡는다. "지금 얼마나 각성/활동 중인가"를 뜻함 — 편차의 최강 신호.
- 심박변이 HRV — 심박 간격의 흔들림(자율신경 상태). 낮으면 교감신경 우위(스트레스).
- 손목 온도 WT — 손목 피부 온도. 말초 혈관 수축/저항(TPR)의 대리 — 추우면 수축→혈압↑.
- 운동강도 METs — 절대적 운동 강도(대사당량). 활동성 급성 부하.
- 활동 active — METs≥3 또는 걸음≥300이면 1(움직이는 중).
- 수면 sleep — 밤(00~06시·22~24시)이면서 비활동이면 1(자는 중, 야간 혈압 저하 반영).
변수별 BP 관계 · 선행연구 레퍼런스
각 워치 변수가 혈압과 어떻게·왜 연결되는지(방향·기전)와 핵심 인용. (다중 에이전트 문헌조사 기반)
심박변이(HRV) — overall autonomic marker vs hypertension
INVERSE and bidirectional. Lower HRV is both a CONSEQUENCE of (cross-sectional, lower in prevalent hypertensives) and a PREDICTOR/ANTECEDENT of higher BP and incident hypertension. Mechanism: reduced vagal tone + relative sympathetic dominance -> impaired baroreflex BP buffering -> sustained sympathetic overactivity -> elevated BP.
- Schroeder EB, Liao D, Chambless LE, et al. 2003, Hypertension (ARIC Study) 42(6):1106-1114. Cohort 11,061 (age 45-64); cross-sectionally HRV (SDNN, RMSSD, R-R) lower in hypertensives across the FULL BP range after adjustment. Prospectively among 7,099 normotensives at baseline, low HRV predicted incident hypertension over ~9 yr: lowest vs highest quartile HR 1.24 (95% CI 1.10-1.40) for SDNN, 1.36 (1.21-1.54) for RMSSD, 1.44 (1.27-1.63) for R-R interval. Establishes the temporal sequence: low HRV precedes hypertension.
- Singh JP, Larson MG, Tsuji H, et al. 1998, Hypertension (Framingham Heart Study) 32(2):293-297. 931 men + 1111 women, first 2h ambulatory ECG. Cross-sectionally HRV significantly lower in hypertensive men and women. Prospectively (4-yr follow-up, 633 men/801 women normotensive at baseline; 119 men/125 women became hypertensive): low LF power predicted new-onset hypertension in MEN (not women); in men LF was a stronger predictor than BMI. Sex-specific direction.
Time-domain HRV (SDNN, RMSSD, pNN50, R-R interval)
INVERSE. SDNN, RMSSD and R-R interval are lower in hypertensives than normotensives and lower values predict incident hypertension. SDNN is the most consistently/significantly reduced; RMSSD/pNN50 reduced but sometimes not statistically significant in smaller studies.
- Schroeder EB et al. 2003, Hypertension (ARIC): crude SDNN, RMSSD, and R-R lower in hypertensives vs normotensives, persisting after adjustment (age, sex, race, center, diabetes, smoking, education, BMI). Lowest-quartile hazard ratios for incident hypertension — SDNN 1.24, RMSSD 1.36, R-R 1.44 (see above).
- ELSA-Brasil — Almeida-Santos / Brazilian Longitudinal Study of Adult Health, 2021, J Hum Hypertens (Nature) — prospective, 7,665 normotensive at baseline, 4-yr follow-up. Low values of HRV indices (SDNN, VLF, LF) significantly associated with increased relative risk of incident hypertension after full adjustment; even normal-BP (not just prehypertensive) subjects with low SDNN/VLF/LF were at increased risk. Conclusion: cardioautonomic disturbance precedes clinical hypertension.
Frequency-domain HRV (LF power, HF power, LF/HF ratio)
INVERSE for HF (lower vagal power -> higher BP/hypertension risk). For LF/HF: hypertensives typically show HIGHER LF and HIGHER LF/HF ratio (relative sympathetic dominance). Low HF predicts incident hypertension; low LF predicts new-onset hypertension in men (Framingham).
- Liao D et al. 1996, Am J Hypertens (ARIC): graded inverse HF-to-incident-hypertension association, lowest HF quartile OR 2.44 (1.15-5.20).
- Singh JP et al. 1998, Hypertension (Framingham): low LF predicted incident hypertension in men.
Baroreflex sensitivity (BRS) — mechanistic link between HRV and BP
INVERSE with BP. Reduced/blunted BRS accompanies and precedes hypertension; impaired baroreflex buffering -> wider BP variability and sustained sympathetic overactivity -> elevated BP. BRS is positively correlated with HRV and inversely with BP variability. Hypertension also RESETS the baroreflex to a higher pressure with reduced sensitivity (reverse causation).
- AHA Hypertension — 'Baroreflex Sensitivity Inversely Correlates With Ambulatory Blood Pressure in Healthy Normotensive Humans' (ahajournals.org HYPERTENSIONAHA.107.090308): lower BRS associated with higher ambulatory BP even within normotensive range.
- Crisafulli A et al. 2022, 'Hypertension depresses arterial baroreflex control of heart rate and cardiac output during rest, exercise, and metaboreflex activation' (PMC9602692): hypertensives have blunted baroreflex control of heart period; baroreflex reset to higher pressure at cost of reduced sensitivity.
HRV as predictor of broader cardiovascular outcomes (context for BP pathway)
INVERSE/predictive but MODEST effect size. Low HRV signals autonomic dysregulation that drives both hypertension and downstream CVD; pooled risk roughly 1.3-1.5x lowest-vs-highest HRV. Supports HRV as an antecedent marker rather than a strong standalone predictor.
- Hillebrand S, Gast KB, de Mutsert R, et al. 2013, Europace 15(5):742-749 — meta-analysis & dose-response meta-regression of HRV and first CV event in populations without known CVD: pooled RR (lowest vs highest SDNN) 1.35 (95% CI 1.10-1.67); LF 1.45 (1.12-1.87); HF 1.32 (0.96-1.81). A 1-SD lower SDNN ~ 32-45% higher CV risk.
- General meta-analytic finding cited across reviews: lower HRV associated with higher all-cause death / CV events in CVD patients (pooled HR for all-cause death ~2.27) — larger effect in diseased populations vs primary-prevention cohorts.
혈중 산소 (SpO2 / blood oxygen saturation)
INVERSE in the clinically dominant range: lower SpO2 / more hypoxia → higher BP. Mechanism: hypoxia activates the carotid-body peripheral chemoreflex → increased sympathetic (muscle sympathetic nerve activity) outflow, RAAS/renin-angiotensin-aldosterone activation, reduced nitric-oxide bioavailability, endothelial dysfunction, oxidative stress and systemic inflammation → vasoconstriction and sustained BP elevation. Intermittent hypoxia (as in OSA) is especially hypertensinogenic, producing nocturnal/resistant hypertension and abnormal BP variability. CAVEAT: relationship is non-linear/context-dependent — very brief or mild acute hypoxia may change BP little, and some chronic high-altitude adaptation or RAAS-inhibition models can lower BP — but the prevailing direction in OSA/desaturation literature is hypoxia→higher BP.
- Peppard et al. 2000, N Engl J Med (Wisconsin Sleep Cohort) — dose-response, BMI-independent: baseline AHI≥15 vs 0 gave ~4-year adjusted OR ≈2.89 (95% CI 1.46-5.64) for incident hypertension; AHI≥15 ~3.2-fold higher odds vs no OSA
- Tamisier et al. 2011, Am J Physiol Heart Circ Physiol — controlled chronic intermittent hypoxia 9h/night x28 nights in healthy young adults: diastolic BP rose 71±1.3→74±1.7 mmHg (P<0.01); MSNA rose 9.94→14.63 bursts/min (P<0.05); demonstrates causal hypoxia→BP+sympathetic link
호흡수 (respiration rate / breathing frequency)
BIDIRECTIONAL / pattern-dependent. (a) SLOW breathing lowers BP: paced breathing ~6 breaths/min raises arterial baroreflex sensitivity and vagal tone while lowering sympathetic and chemoreflex drive → reduced SBP/DBP. (b) HIGH respiratory rate signals sympathetic activation / acute physiological stress and is an independent predictor of deterioration and mortality (often stronger than BP or pulse), implying association with adverse cardiovascular state. (c) MECHANISTIC coupling: respiration generates short-term BP oscillations via respiratory sinus arrhythmia and the baroreflex — BP falls during inhalation and rises during exhalation; faster breathing produces smaller respiratory BP swings. Net: slow/regular breathing ↔ lower BP & higher baroreflex sensitivity; fast/irregular breathing ↔ sympathetic activation & higher risk.
- Joseph et al. 2005, Hypertension 46:714-718 — slow breathing 6/min in essential hypertensives: SBP 149.7±3.7→141.1±4 (P<0.05), DBP 82.7±3→77.8±3.7 mmHg (P<0.01); baroreflex sensitivity rose 5.8±0.7→10.3±2.0 ms/mmHg (hypertensives) and 10.9→16.0 (controls); reduced sympathetic/chemoreflex activation
- Mahtani et al. 2012, J Hum Hypertens (RESPeRATE device meta-analysis, 8 trials, n=494) — SBP −3.67 mmHg (95% CI −5.99 to −1.39, P=0.002), DBP −2.51 mmHg (−4.15 to −0.87, P=0.003); BUT effect null after excluding manufacturer-linked trials
심박수 (Heart Rate, HR) — general/instantaneous
Context-dependent (this is the critical modeling caveat). INVERSE at rest on a beat-to-beat basis via the arterial baroreflex: an acute BP rise activates baroreceptors -> reflex parasympathetic activation + sympathetic inhibition -> HR, contractility, stroke volume, and vascular resistance fall (negative feedback keeping BP constant); a BP fall does the opposite. POSITIVE/parallel during exercise because the baroreflex resets upward intensity-dependently and central command + exercise pressor reflex drive HR and BP up together to increase CO. POSITIVE across individuals chronically: higher tonic HR co-occurs with higher BP because both are downstream of shared sympathetic overdrive.
- Baroreflex Sensitivity Inversely Correlates With Ambulatory Blood Pressure in Healthy Normotensive Humans, Hypertension (AHA, ahajournals.org/doi/10.1161/HYPERTENSIONAHA.107.090308) — baroreflex enforces inverse SBP-HR relation; baroreflex sensitivity inversely correlates with ambulatory BP in normotensives.
- Human/animal baroreflex reviews (PMC9602692; Fadel et al., PMC3253263 'Human Investigations into the Arterial and Cardiopulmonary Baroreflexes during Exercise') — during exercise the arterial baroreflex resets in direct proportion to intensity, allowing parallel rises in BP and HR to raise cardiac output; exercise pressor/metaboreflex raises BP via increased HR, contractility (Emax), and CO.
휴식기 심박수 (Resting Heart Rate, RHR)
POSITIVE and independent as a long-term cross-individual predictor: elevated RHR precedes and predicts incident hypertension and higher BP. Mechanism = chronic sympathetic overactivity that raises both HR and BP (and hematocrit/metabolic abnormalities). RHR >80 bpm flagged as marker of pronounced sympathetic overdrive with adverse outcomes. (Note: this long-term positive risk association is distinct from the acute within-person baroreflex inverse relation.)
- Wang et al. (or cohort authors), 'Association of Resting Heart Rate With Blood Pressure and Incident Hypertension Over 30 Years in Black and White Adults: The CARDIA Study,' Hypertension, Sept 2020 (AHA, HYPERTENSIONAHA.120.15233; PMC7430042) — n=3888 (1615 men, 2273 women) followed ~30 yrs; HR per +10 bpm RHR = 1.47 (1.23-1.75) Black men, 1.51 (1.28-1.78) White men, 1.48 (1.26-1.73) White women, 1.02 (0.89-1.17, NS) Black women.
- Dose-response meta-analysis of cohort studies (PMC7478507) — linear positive association; ~9% higher incident-HTN RR per +10 bpm; RHR ~100 vs 55.5 bpm = RR 1.58 (1.32-1.88); no significant association in women subgroup.
걷기 심박수 (Walking / activity-elevated Heart Rate) — incl. heart-rate reserve & HR-per-step
Two layers. (a) WITHIN an activity bout: POSITIVE co-movement — walking raises HR and BP together (baroreflex reset + exercise pressor reflex). (b) As a TRAIT predictor of hypertension, the normalized response matters and direction can flip: HIGHER heart-rate reserve / good chronotropic competence is PROTECTIVE against incident HTN (better fitness, lower resting sympathetic tone), whereas a STEEP BP-rise-per-HR-increment slope and an exaggerated exercise BP response predict FUTURE hypertension. So absolute walking HR co-moves positively with BP acutely, but a blunted HR reserve or a steep BP/HR slope flags higher hypertension risk.
- Sharashova et al., 'Association of exercise heart rate response and incidence of hypertension in men,' (PubMed 24974261) — higher HR reserve protective: HR 0.84 (95% CI 0.74-0.95) for highest vs lowest quartile of HR reserve for incident hypertension in men.
- Singh JP et al., 'Blood Pressure Response to Heart Rate During Exercise Test and Risk of Future Hypertension,' Hypertension 2002 (AHA hy0302.105777; PubMed 11897759) — Framingham-related; n~1033 normotensive men, mean age ~43; steeper BP-response-to-HR slope (and percentile curves via 3rd-order polynomial regression) predicts future hypertension; exaggerated exercise BP response RR ~3.8 (2.3-6.1) for incident HTN.
Wearable wrist HR / PPG as a BP estimator (cross-cut, supports HR variables above)
HR alone is a WEAK standalone BP proxy; predictive value comes from combining HR with PPG waveform morphology, PAT, and demographics. Best cuffless wristbands report SBP/DBP mean errors near +/-2.5 to 3.7 mmHg seated, but accuracy degrades by posture (SD up to ~12.5 mmHg standing) and overall accuracy remains regulatorily uncertain. Implication for a BP-estimation model: include resting HR (level/risk), activity-context HR (co-moves with BP), HR reserve / HR-per-step (fitness/autonomic responsiveness), plus PPG-derived morphology.
- Evaluation of a novel cuffless PPG-based wristband for measuring blood pressure per regulatory standards, Eur Heart J Digital Health 2024 (academic.oup.com/ehjdh/article/5/3/335; PMC11104472) — algorithm errors ~+/-3.7 (SD 4.4) mmHg SBP and +/-2.5 (SD 3.7) mmHg DBP; requires cuff initialization + ~100 features.
- Frontiers Med Technol 2024 (10.3389/fmedt.2024.1464473) PPG cuffless bracelet — SBP mean diff 0.5 (SD 7.8) mmHg sitting, -2.4 (10.1) supine, -0.6 (12.5) standing; accuracy posture-dependent.
보행 속도 (Gait speed / usual walking speed)
Inverse / protective: faster gait speed predicts LOWER incident hypertension and LOWER CVD/all-cause mortality; slower gait associates with HIGHER arterial stiffness and worse CV outcomes. Direction: faster gait → better BP/CV profile. Caveat: largely observational; reverse causation (vascular disease slows gait) is plausible.
- Veronese N et al. 2018, J Am Med Dir Assoc (JAMDA) — meta-analysis, 45 studies / ~101,945 adults (mean age 72.2y, median FU 5.4y): each 0.1 m/s SLOWER gait = +12% all-cause mortality (HR 1.12, 95%CI 1.09-1.14) and +8% CVD (HR 1.08, 95%CI 1.03-1.13).
- LaCroix AZ / WHI 2020, Hypertension (HYPERTENSIONAHA.120.15839) — 83,435 postmenopausal women, mean 11y FU, 38,230 incident treated-HTN cases: faster usual walking speed inversely associated with incident treated hypertension; HRs by pace ~1.00 (<2 mph, ref), 0.95, 0.86, 0.79 (>4 mph); significant after adjusting for baseline BP and walking volume.
보폭 (Stride length / step length)
Inverse / protective: SHORTER stride length predicts adverse clinical events and disability; reduced step length is a frailty feature linked to higher CV risk. Direct BP/stiffness associations are sparse and mostly mediated through gait speed and frailty rather than a standalone BP mechanism.
- Stride Length Predicts Adverse Clinical Events — Systematic Review & Meta-analysis 2021, PMC8235531: short stride length predicted major adverse events (pooled OR 1.36) and physical disability (OR 1.26) in older adults.
- Frailty/gait studies (Frontiers Med 2026, 3389/fmed.2026.1785926) — frail older adults show shorter step length, prolonged double-support, and reduced gait speed vs robust; frailty phenotype associates with higher CV risk.
이중 지지 시간 (Double-support time / gait variability)
Positive (longer double-support / higher variability = worse): hypertensive individuals walk slower with LONGER double-support time and GREATER gait variability vs normotensives. Increased variability marks frailty/fall risk and tracks with CV disease burden. Direction: longer double-support/higher variability → higher BP/CV-risk phenotype, but mechanism is indirect (shared vascular-neural aging).
- Determining characteristics of gait variability at preferred speed in hypertensive vs normotensive participants 2023, J Clin Hypertens (S2213-3981(23)00131-8) — hypertensive participants: slower speed, longer double-support time, and significantly greater gait variability than normotensives.
- Medical, Sensorimotor & Cognitive Factors Associated with Gait Variability — longitudinal population study, PMC6305368: baseline CVD increased rate of change of step-length variability; weaker quadriceps increased step-width variability change.
보행 비대칭성 / 안정성 (Gait asymmetry / walking steadiness/stability)
Indirect, disease-mediated: gait asymmetry/instability is elevated in PAD (correlates with ankle-brachial index and ABI asymmetry — a peripheral vascular/atherosclerosis readout) and in stroke survivors. No clean direct BP-magnitude relationship; rather, vascular disease (atherosclerosis/PAD, cerebrovascular events from hypertension) produces asymmetry/instability. Direction: greater asymmetry/lower steadiness → greater underlying vascular/CV disease burden.
- Effect of PAD & intermittent claudication on gait regularity and symmetry 2022, J Biomech (PubMed 35759975 / S0021-9290(22)00248-2): gait asymmetry increased after claudication onset; ML asymmetry and vertical stride regularity correlated with ankle-brachial index (ABI) and ABI asymmetry.
- Gait pattern in patients with PAD 2018, BMC Geriatrics (PMC5819174): reduced gait speed, step/stride regularity vs controls; affected limb longer swing, shorter stance.
계단 올라가기 속도 (Stair-climbing speed / capacity)
Inverse / protective as a fitness marker: greater stair-climbing activity/capacity associates with LOWER CVD and all-cause mortality and better CV risk-factor profile. Acutely, stair climbing raises HR, stroke volume, and systolic BP (training stimulus); chronically improves fitness and lowers risk. Direction: higher stair-climbing capacity → better CV/BP profile. Caveat: reverse causation (fitter people climb more).
- Paddock S et al. 2024, Eur J Prev Cardiol (zwae175.405) — systematic review/meta-analysis, 9 studies / 480,520 participants (pooled 5 studies, 455,649): stair climbing associated with 24% lower all-cause mortality and 39% lower CVD mortality; pooled CV-mortality RR 0.65 (95%CI 0.50-0.83).
- Daily stair climbing, disease susceptibility & ASCVD — prospective cohort 2023 (S0021-9150(23)05221-8): climbing >5 flights (~50 steps)/day associated with lower ASCVD risk independent of genetic susceptibility.
일어서기 시간 / 좌식 시간 (Standing/sitting & sedentary time)
Positive (sedentary = worse): MORE sitting / longer uninterrupted sitting acutely RAISES blood pressure and aortic PWV (arterial stiffness) and is dose-dependently associated with incident hypertension and CVD. Replacing sitting with movement lowers BP, but merely STANDING (without ambulating) does NOT reliably lower BP and excessive standing may harm. Direction: more sedentary time → higher BP/stiffness/CV risk; light activity (not just standing) → lower BP.
- Barone Gibbs B et al. 2024, Circulation (RESET-BP RCT, CIRCULATIONAHA.123.068564) — reducing sedentary behavior lowered BP in desk workers; companion analysis (WVU 2024) — increasing standing ~1 h/day for 3 months did NOT reduce BP or arterial stiffness.
- Prolonged Sitting Induces Elevated BP in Healthy Young Men — RCT crossover 2024, PMC10981358; and Single Bout of Prolonged Sitting Augments Very Short-Term BP Variability 2024, Am J Hypertens (PMC11322278) — acute sitting raises BP / BP variability.
손목 온도 (wrist / distal skin temperature)
INVERSE / negative, mechanistically driven by peripheral vasoconstriction–vasodilation (total peripheral resistance). Higher wrist/distal skin temperature -> lower SBP/DBP. Across the day the BP rhythm is approximately a mirror image of the distal skin temperature (and DPG) rhythm. Low DAYTIME distal skin blood flow (cooler distal temp by day) predicts blunted nocturnal BP dipping (non-dipping), so the wrist-temperature circadian profile can screen dipper vs non-dipper status. Mechanism: nocturnal parasympathetic activation + supine position -> distal vasodilation -> heat loss + BP fall.
- Wang et al. 2022, Hypertension (AHA) 'Inverse Association of Skin Temperature With Ambulatory Blood Pressure...' (Nara Medical Univ.) — n=584 older adults, 48h ABPM + skin temp (30,711 daytime / 17,382 nighttime readings, Oct–Mar). Distal (wrist+ankle) and proximal (abdomen) skin temp inversely associated with daytime SBP: −4.27 mmHg (95% CI −4.58 to −3.96) and −2.74 mmHg per SD of skin temp; associations also significant at night; skin temp mediated BP responses to ambient temperature (path analysis). PMID 35574922; DOI 10.1161/HYPERTENSIONAHA.122.19190
- Blázquez/Sarabia/Madrid et al. 2012, Chronobiology International 29(6):747 'Wrist Skin Temperature, Motor Activity, and Body Position as Determinants of the Circadian Pattern of Blood Pressure' — wrist temp negatively correlated with SBP/DBP; 78.6% concordance between observed BP dips and BP pattern predicted from wrist-temperature rhythm. PMID 22734575
일광 시간 (daylight / light exposure)
DUAL / direction depends on timing. (1) Acute photobiology: solar UV-A mobilizes cutaneous nitric-oxide stores -> systemic vasodilation -> LOWER SBP (independent of temperature). (2) Circadian: adequate DAYTIME / MORNING bright light reinforces the nocturnal BP dip and supports lower daytime BP; LACK of morning light and LIGHT-AT-NIGHT blunt dipping, raise nighttime BP/morning surge and CV risk. Seasonality measured as DAYLIGHT HOURS is an independent predictor: nighttime SBP positively related to daylight hours and morning BP surge negatively related to daylight hours (i.e., long photoperiod/summer -> higher nighttime SBP, smaller morning surge), distinct from the temperature effect on daytime SBP. Net wintertime (short daylight + cold) elevation of daytime BP.
- Weller/Feelisch et al. 2020, JAHA 9(5):e013837 'Does Incident Solar Ultraviolet Radiation Lower Blood Pressure?' — large N. American chronic hemodialysis cohort; higher incident solar UV associated with lower SBP after adjusting for ambient temperature; supports UV->cutaneous NO->lower BP mechanism. PMC7335547
- Liedtke/… 2023, Scientific Reports 13 (Nature) 'Low-dose daylight exposure induces nitric oxide release and maintains cell viability in vitro' — mechanistic support that daylight-level exposure triggers NO release. DOI 10.1038/s41598-023-43653-2
수면 (sleep duration / quality / efficiency)
Generally INVERSE for quantity/quality vs BP and dipping: SHORT or POOR sleep -> HIGHER BP, higher incident hypertension, and BLUNTED nocturnal dipping / non-dipping (sympathetic over-activity, reduced parasympathetic tone at night). Better subjective sleep quality -> more likely dipper. Sleep also defines the morning-surge/dip windows. Evidence is strongest for short duration and subjective quality; some OBJECTIVE measures (actigraphic duration/efficiency) show weaker/null links to non-dipping (CARDIA). OSA (a sleep disorder) strongly raises non-dipping prevalence.
- Lo, Woo et al. 2018, Journal of Clinical Hypertension 20(3):592 'Subjective sleep quality, blood pressure, and hypertension: a meta-analysis' — 29 articles/45,041 patients; poor sleep quality associated with hypertension (OR 1.48, P=.01); highest sleep-disturbance tertile OR 1.44 (1.00–1.64); poor sleepers higher SBP (MD 4.37) and DBP (MD 1.25) (NS). Notes dippers have better subjective sleep quality than non-dippers. PMID 29457339
- Gangwisch et al. 2006, Hypertension (NHANES I, n≈4,810, 8–10y) — sleep ≤5 h/night raised incident hypertension (adjusted HR 1.32, 95% CI 1.02–1.71 vs 7–8h); ages 32–59 with ≤5h ~60% more likely to develop hypertension.
걸음 (Step count / steps per day)
Inverse: higher daily step volume associated with lower SBP, DBP, MAP and pulse pressure, and lower MACE in hypertensives. Mechanism — chronic adaptation lowering peripheral resistance, improved endothelial/NO vasodilation, better baroreflex, reduced sympathetic tone; dose-dependent up to ~8,000–10,000 steps.
- del Pozo Cruz et al. 2024, Eur J Prev Cardiol — UK Biobank hypertensives (n≈32,192): each +1,000 steps/day vs ~2,300 baseline ~17% lower MACE risk.
- Chen et al. 2024, JMIR — higher wearable daily steps associated with lower SBP/DBP/MAP/pulse pressure in older adults; greater step variability linked to higher BP.
활동 에너지 (Active energy expenditure / MVPA)
Inverse, dose-responsive: greater active energy / MVPA lowers resting BP and incident hypertension via chronic vascular remodeling + repeated acute post-exercise hypotension. Caveat: wrist active-kcal carries large error (MAPE often >20%).
- Stamatakis et al. (ProPASS) 2025, Circulation (n=14,761) — reallocating time toward exercise-like activity yields clinically meaningful BP reductions (≥2 mmHg SBP).
- Liu et al. 2017, Hypertension — ~6% lower hypertension risk per 10 MET-h/week of leisure-time activity.
휴식 에너지 (Resting / basal energy, RMR)
Positive (opposite to activity metrics): higher resting/basal metabolic rate associated with higher BP independent of body size — elevated resting rate-pressure product, greater sympathetic activity, lower insulin sensitivity.
- Weyer et al. 2000, Hypertension — obese hypertensives show significantly higher RMR vs BMI-matched normotensives.
- MR study 2023, Sci Rep — genetically higher BMR causally raises heart failure / aortic aneurysm risk (adverse CV direction).
절대적 운동강도 METs (Exercise intensity)
Inverse and intensity-graded: higher intensity lowers BP more per unit time; reductions scale with MET-minutes. Stronger acute post-exercise hypotension + chronic drop in peripheral resistance/arterial stiffness; HIIT especially effective.
- Network meta-analysis 2023, Sci Rep — HIIT ~690 MET-min/week most effective for SBP (−5.07 mmHg); ~830 MET-min/week best for DBP (−4.42 mmHg).
- Pescatello et al. (ACSM) — regular aerobic exercise lowers BP ~5–7 mmHg in hypertensives.
운동하기 시간 (Exercise minutes / duration)
Inverse: accumulating exercise duration (~150 min/week) lowers resting SBP/DBP. Acute — each 20–60 min bout >40% VO2peak triggers post-exercise hypotension up to ~24 h; chronic — sustained ambulatory BP reduction.
- Resistant-HTN RCT 2021 (PMC8340008) — 12 wk of 3×40-min/week reduced 24-h and daytime ambulatory SBP/DBP.
- Halliwill 2001, post-exercise hypotension review — PEH after submaximal dynamic exercise persists for hours (up to ~13 h).
걷기+달리기 거리 (Walking + running distance)
Inverse, energy-driven: greater distance lowers incident hypertension and SBP/DBP; at matched energy expenditure walking and running give comparable benefit.
- Williams & Thompson 2013, ATVB (n≈49,000) — running cut incident hypertension 4.2% and walking 7.2% per MET-h/day; comparable at equal energy.
- Walking intervention 2018 (PMC6119598) — 6-month guided walking lowered SBP up to −21.3 mmHg in those with baseline SBP >160.
오른 층수 (Flights climbed / stair climbing)
Inverse: habitual stair climbing lowers SBP and arterial stiffness, increases leg strength; more flights/day associated with lower CVD risk. Brief vigorous weight-bearing load reduces arterial stiffness (key SBP mediator).
- Hong et al. 2018 (Korean postmenopausal stage-2 HTN) — 12-wk stair-climbing reduced arterial stiffness and BP; stiffness drop partly explained SBP improvement.
- Tulane Mediterranean cohort — climbing >5 flights/day associated with ~20% lower ASCVD risk.
편차
편차 모델 — 워치(급성 생리)
E0를 고정하고 편차 r = 실측 − E0만 모델한다(잘 보정된 수준·위상을 다시 흔들지 않기 위해 residual mode). 입력은 워치 8개(급성 생리)뿐 — 과거 BP추정변수(전날·과거일 동시각·과거일 평균)는 제외했다. 이유: 그 lag들은 이미 E0(수준+위상)가 담고 있는 사람 level과 대체로 중복이라, 빼도 성능이 거의 유지되면서 구조가 훨씬 깔끔해진다(자기 추정치에 되먹임하지 않음).
워치가 잡는 것. 워치는 동시점(급성) 신호다 — 심박·ΔHR(각성)·활동이 오르면 그 순간 BP도 오른다. 특히 세션 DBP(이완기)와 일평균을 E0 위에서 더 깎는다. 어떤 워치 변수가 얼마나 기여하는지는 아래 SHAP 해석에서.
zoo
전 모형 zoo — 20개, train/valid/test
같은 편차 과제(워치변수만)에 회귀·스펙트럴·웨이블릿·트리·GBM·딥러닝을 single·multi로 올렸다(최종=E0+r̂). 셀 = MAE + BHS등급. days2-7, 검증·테스트 각 24명.
| 모형 | 검증 valid (716세션) | 테스트 test (697세션) | ||||||
|---|---|---|---|---|---|---|---|---|
| 세SBP | 세DBP | 일SBP | 일DBP | 세SBP | 세DBP | 일SBP | 일DBP | |
| 수준+커널 E0 | 6.17C | 5.12B | 3.66A | 2.50A | 5.89B | 4.83B | 2.41A | 2.35A |
| Ê=E0+워치편차(ET) | 6.24C | 5.12B | 3.76A | 2.44A | 5.82C | 4.76A | 2.50A | 2.31A |
| 회귀 Ridge | 6.10C | 5.01B | 3.59A | 2.29A | 5.77B | 4.76B | 2.40A | 2.28A |
| 회귀 Lasso | 6.10C | 5.01B | 3.58A | 2.30A | 5.79B | 4.75A | 2.43A | 2.27A |
| 회귀 ElasticNet | 6.10C | 5.01B | 3.58A | 2.30A | 5.78B | 4.75B | 2.42A | 2.28A |
| 회귀 MultiTaskElasticNet | 6.10C | 5.01B | 3.58A | 2.30A | 5.78B | 4.75B | 2.42A | 2.28A |
| 스펙트럴 Harmonic+Ridge | 6.13C | 5.03B | 3.60A | 2.29A | 5.76B | 4.68A | 2.38A | 2.29A |
| 스펙트럴 Harmonic+MTEN | 6.12C | 5.02B | 3.60A | 2.30A | 5.77B | 4.69A | 2.40A | 2.28A |
| 트리 RandomForest | 6.46C | 5.24B | 3.95A | 2.54A | 5.90C | 4.82A | 2.36A | 2.38A |
| 트리 ExtraTrees | 6.25C | 5.15B | 3.78A | 2.47A | 5.83B | 4.76A | 2.46A | 2.33A |
| 트리 ExtraTrees(multi) | 6.22C | 5.17B | 3.75A | 2.48A | 5.83B | 4.75A | 2.51A | 2.30A |
| GBM XGBoost | 6.43C | 5.23B | 3.93A | 2.55A | 5.99C | 4.76A | 2.39A | 2.29A |
| GBM LightGBM | 6.63C | 5.30B | 4.05A | 2.58A | 6.10C | 4.87A | 2.24A | 2.33A |
| GBM CatBoost | 6.30C | 5.16B | 3.82A | 2.50A | 5.79B | 4.70A | 2.31A | 2.22A |
| GBM XGBoost(multi) | 6.43C | 5.20B | 3.91A | 2.51A | 5.91C | 4.78A | 2.35A | 2.20A |
| GBM CatBoost(multi) | 6.26C | 5.13B | 3.77A | 2.50A | 5.76B | 4.68B | 2.29A | 2.19A |
| 딥러닝 MLP | 6.37C | 5.06B | 3.83A | 2.40A | 5.84B | 4.71A | 2.64A | 2.29A |
| 딥러닝 MLP(multi) | 6.08C | 5.04B | 3.60A | 2.45A | 5.81C | 4.86B | 2.49A | 2.39A |
전 모형 검증 MAE(계열색). 선형(회귀)·스펙트럴이 낮고 안정적, 트리·GBM은 학습을 외워 held-out에서 높다. 점선=E0 기준.
🏆 챔피언 = 회귀 Ridge (≈ ElasticNet·Harmonic+Ridge, 선형 shrinkage). 검증 세션 SBP 6.10 C·DBP 5.01 B, 일평균 SBP 3.59 A·DBP 2.29 A. 테스트 세션 SBP 5.77 B·DBP 4.76 B, 일평균 SBP 2.40 A·DBP 2.28 A — 워치+구조만으로 일평균 Grade A. (Harmonic+Ridge면 테스트 세션 DBP 4.68 A)
해석
최종모형 해석 — SHAP + 설명변수 시차
챔피언 회귀 Ridge(편차 모델, 워치변수만)가 무엇을 보고 예측하는지 SHAP으로, 남은 잔차의 구조를 함께 분석했다.
워치 변수 SHAP 중요도(편차 r에 대한 기여). SBP는 ΔHR(0.86)·활동·HRV, DBP는 심박수 HR(0.97)·ΔHR·활동이 주력. 즉 심혈관 각성(HR·ΔHR)이 편차의 핵심. +/− = 방향, 색 = 신호 계열(심박·활동·체온), 진한막대=SBP·연한막대=DBP.
읽기. 편차를 줄이는 워치 신호는 ΔHR·심박수(각성/활동 → BP↑)가 지배적이고, 특히 이완기(DBP)에 HR 기여가 크다. 손목온도·수면은 작은 보조.
워치 기여 + 잔차 구조. 왼쪽: 워치의 편차 총기여 — SBP보다 DBP(이완기)에 더 크게 기여. 오른쪽: 잔차 자기상관이 세션 lag1에서 0.09로 낮음(이후 ≈0), 워치 동시점 상관 약 0.13.
요약. 모형은 동시점(lag 0) 워치로 급성 편차를 잡는다(심박·각성 중심, DBP에 유리). E0+편차 이후 세션 잔차 자기상관이 0.09로 낮아 그날의 출렁임은 본질적으로 예측이 어렵다 — 일평균이 세션보다 정확한 이유이자 세션 SBP가 Grade B에 머무는 구조적 한계와 일치.
관계
설명변수 ↔ 타깃 산점도 — 무엇이 혈압과 함께 움직이나
개별 설명변수(수준 E0 + 워치 8종)와 세션 혈압의 관계를 전체 3,549세션 산점도로 본다. 주황=OLS 회귀선, r=Pearson 상관, ***p<.001.
설명변수 생성 방법 — 원천 → 파생식 (어떻게 만들었나)
산점도의 9개 변수를 각각 이렇게 만들었다. 원신호 점검·생리 기전·선행연구 레퍼런스 상세는 §2-5, 수준 E0 상세는 §2-1~2-4.
| 설명변수 | 원천 | 생성 방법 (파생식) |
|---|---|---|
| 수준 E0 | day1 실측 + KNHANES 외부앵커 L + 군집(연령×성별) 일주기 | E0 = lvl + α·(day1 편차의 역시간×위상 커널가중) + (1−α)·군집 위상곡선. lvl = (n·day1평균 + 2·L)/(n+2), α = Σw/(Σw+5) |
| 심박수 HR | last_심박수 | 세션의 마지막 심박수 값 |
| 커널가중 심박 kwHR | kw_심박수 | 그 시점 주변 심박을 시간커널로 가중평균(순간 튐 완화) |
| 안정대비 심박상승 dHR | HR · 개인 HR 분포 | HR − 안정심박(개인 HR 하위 10% 분위) |
| 심박변이 HRV | last_심박 변이 | 세션 마지막 HRV (자율신경 상태) |
| 손목온도 WT | last_손목 온도 | 세션 마지막 손목 피부온도 (말초저항 대리) |
| 대사당량 METs | last_절대적 운동 강도 | 세션 마지막 METs (절대 운동강도) |
| 활동 active | METs · last_걸음 | METs≥3 또는 걸음≥300 → 1 |
| 수면 sleep | 시각 tod · active | 야간(06시 이전 또는 22시 이후) & 비활동 → 1 |
공통 처리. 결측은 train 데이터의 중앙값으로 대치(train·role=train에서만 적합 — 누수 방지). 과거 BP 추정치 파생변수(pmean·ptyp·lag1 등)는 모델에서 제외. 워치 8종은 raw BP가 아니라 편차 r = BP − E0를 표적한다.
| 설명변수 | SBP r | DBP r | 방향 · 해석 |
|---|---|---|---|
| 수준 E0 (설문·외부·day1) | +0.81 | +0.72 | 강한 직선 — "누구인가(수준)" |
| 손목온도 WT | −0.13 | −0.14 | 온도↑ → BP↓ (혈관확장) |
| 심박수 HR | +0.07 | +0.15 | 심박↑ → BP↑ |
| 안정대비 심박상승 dHR | +0.08 | +0.06 | 운동부하↑ → BP↑ |
| 대사당량 METs | +0.09 | +0.04 | 활동↑ → BP↑ |
| 활동량 active (0/1) | +0.09 | +0.04 | 활동중 → BP↑ |
| 심박변이 HRV | −0.05 | −0.05 | HRV↑ → BP↓ |
| 커널가중 심박 kwHR | +0.05 | +0.01 | 약함 |
| 수면 sleep (0/1) | −0.02 | −0.04 | 매우 약함 |
세션 SBP vs 개별 설명변수(|r| 내림차순 3×3). 수준 E0만 뚜렷한 직선, 워치는 흩어짐. 활동·수면은 0/1 이진이라 두 세로 밴드.
세션 DBP vs 개별 설명변수. DBP는 심박수 HR(r=0.15)이 워치 중 가장 강하다.
✅ 수준 E0만 압도적(r 0.72~0.81, 거의 직선). 모델의 환자간(누구인가) 능력이 이 한 장에 그대로 드러난다.
🌀 워치는 전부 약함(|r|≤0.15). 단 N=3,549이라 HR·WT·dHR·METs는 통계적으로 유의하고 방향도 생리학적으로 맞다(심박·활동↑→BP↑, 손목온도·HRV↑→BP↓). "도긴개긴"이 아니라 크기가 작은 진짜 신호.
워치가 약해 보이는 건 raw BP가 아니라 편차 r=BP−E0(그 사람 안 변동)를 표적하기 때문이기도 하다. 이 그림이 곧 다음 절 환자간/환자내 분해의 원인이다.
분해
환자간 vs 환자내 — 오차는 어디서 오나
세션 오차를 두 축으로 분해했다 — 환자간(사람들 사이 = 그 사람 평소 수준을 맞추나)과 환자내(한 사람 안 = 그날그날 시간 변동을 맞추나). 개인 out-of-fold 챔피언 기준.
| 구분 | 세션 SBP | 세션 DBP |
|---|---|---|
| 전체 세션 MAE | 6.45 | 5.11 |
| 환자간(수준): 전체평균 → 모델 | 9.70 → 3.53 | 5.65 → 2.44 |
| 환자간 순위상관 r | 0.93 | 0.90 |
| 환자내(변동): 목표 → 모델오차 | 5.67 → 5.46 | 4.41 → 4.37 |
| 환자내 설명 비율 | 4% | 1% |
세션 오차의 환자간/환자내 분해(왼쪽 SBP·오른쪽 DBP). 환자간(파랑)=개인 수준 오차(전체평균 → 모델), 환자내(빨강)=한 사람 안 시간변동(목표 변동 → 모델 오차). 모델은 환자간을 크게 줄이지만(r 약 0.9) 환자내는 거의 못 줄인다.
✅ 환자간(누구인가): 아주 잘한다. 개인 수준 오차를 SBP 9.70 → 3.53으로 크게 줄이고, 환자 순위를 r=0.93으로 정확히 맞춘다. 외부앵커+day1 보정이 "평소 혈압"을 잘 잡는다.
🌀 환자내(그날 변동): 거의 못 한다. 한 사람 안의 세션 변동(SBP 5.67)을 단 4%(DBP 1%)만 설명. 집계 워치엔 순간 변동 정보가 거의 없다.
이게 "일평균 A · 세션 B"의 근본 이유다. 세션 MAE(6.45)는 못 잡는 환자내 변동(5.46)이 지배한다. 일평균은 이 변동을 평균으로 상쇄해 잘 맞추는 환자간(수준)만 남으므로 Grade A. 세션 Grade A(≤5)엔 환자내 변동을 봐야 하는데, 그건 raw PPG·PTT가 필요하다.
진단
과적합 — 단순 모형이 일반화한다
학습 vs 검증 세션 SBP MAE. 대각선 아래로 멀수록 overfit. 트리·GBM(LightGBM 학습 5.02→검증 6.63)은 학습을 외우고 held-out에서 붕괴. 회귀·스펙트럴은 대각선 근처(학습 6.77→검증 6.10)로 안정.
🌀 트리·GBM은 사람 패턴을 memorize. 학습(in-sample)에선 최고(LightGBM 일DBP 1.59)지만 신규 사람(valid·test)에선 세션 등급이 떨어진다. 워치만 쓰는 지금도 동일한 패턴 — 선택은 valid 기준으로, 트리·GBM 제외하고 안정적인 선형(Ridge)을 챔피언으로.
∎
결론 — 그리고 분단위 곡선
🏁 추정 최종. E0(수준+위상)가 일평균을 BHS A로 잡고, 워치 편차가 세션 DBP·일평균을 더 깎는다 — 과거 BP추정변수 없이도. 챔피언 회귀 Ridge — 테스트 세션 SBP 5.77 B / DBP 4.76 B, 일평균 SBP 2.40 A / DBP 2.28 A. (검증 일평균 SBP 3.59 A / DBP 2.29 A.)
분단위 추정모형. 모형은 임의 분에서 E0와 r̂을 산출해 분단위 혈압 곡선을 그린다(세션·일평균은 그 집계). days2-7 전부를 추정에 활용하며, 별도의 예측 구간은 두지 않는다.
🧱 세션 SBP는 여전히 Grade B. 워치+구조만으론 순간(세션) SBP가 약 5.8(B)에서 멈춘다 — 위상이 집단 평균 리듬이라 그날 급변을 못 담고, 잔차에 시간 구조가 거의 없기 때문(SHAP 참조). 세션 Grade A엔 raw PPG 파형·PTT·주기 재보정이 필요하다(논문 한계 분석).
120명 · 세션평균 하이브리드 보간 · 사람 60:20:20 · days2-7 전부 추정(예측구간 미분리) · 수준(KNHANES+설문+day1)+위상+편차(워치만) · 전 모형 single·multi · MAE+BHS · 세션·일평균