03๋ฐฐ๊ฒฝ ๊ธฐ์ Background
3.1 ์ ์ ๊ณผ ์ง๋จ์ ๊ทผ๋ณธ์ ํ๊ณ3.1 Fundamental Limitations of Psychiatric Diagnosis
ํ์ฌ ํ์ค์ ์ ์ ๊ณผ ์ง๋จ์ ๋ค์ ๋ ๊ฐ์ง ๊ทผ๋ณธ์ ํ๊ณ๋ฅผ ๊ฐ์ง๋ค:
Current standard psychiatric diagnosis is fundamentally limited in two ways:
- ์ฃผ๊ด์ ์๊ธฐ๋ณด๊ณ ์์กด์ฑ: PHQ-9, GAD-7, ADHD-RS ๋ฑ์ ์ฒ๋๋ ๋ชจ๋ ํ์ ์๊ธฐ๋ณด๊ณ ์ ์์กดํ๋ค. ์ฌํ์ ๊ธฐ๋ ํธํฅ, ์๋ ์ํฐ๋ฏธ์, ๊ฐ๋ฉด ์ฆ์ ๋ฑ์ด ์ง๋จ์ ์๊ณกํ๋ค.Reliance on subjective self-report: instruments such as the PHQ-9, GAD-7, and ADHD-RS all rely on patient self-report. Social desirability bias, alexithymia, masked symptoms, and similar factors distort diagnosis.
- ๋จ์ผ ๋ฐ์ดํฐ ์์ค: ๊ฐ๊ด์ ๋ฐ์ด์ค๋ง์ปค(qEEG, HRV, ERP)๋ ๋ณ๋ ํ๊ฐ์ ๊ทธ์น๊ณ , ์๊ธฐ๋ณด๊ณ ์ ํตํฉ๋์ง ์๋๋ค. ๋ค์ค ์ ๋ณด๋ฅผ ์์๊ฐ ์ง๊ด์ ์์กดํด ํฉ์ฐํ๋ ๋นํ์คํ ๊ณผ์ .Single-source data: objective biomarkers (qEEG, HRV, ERP) typically stop at standalone evaluation and are not integrated with self-report; multi-source data is fused only via clinician intuition in a non-standardized fashion.
๊ฒฐ๊ณผ์ ์ผ๋ก ์์์์ ADHD, ์ฐ์ธ์ฆ, ๋ถ์์ฅ์ ๋ฑ์ ์ง๋จ ์ผ์น๋(inter-rater reliability)๋ ฮบ = 0.40-0.60 ์์ค์ ๋จธ๋ฌผ๊ณ ์์ผ๋ฉฐ (Lieblich et al. 2015), ์ด๋ ์ํ์ ๋ค๋ฅธ ๋ถ์ผ์ ๋นํด ํ์ ํ ๋ฎ๋ค.
Consequently, inter-rater reliability for clinical diagnoses of ADHD, depression, anxiety disorders, and similar conditions remains at ฮบ = 0.40โ0.60 (Lieblich et al. 2015), which is markedly lower than in other fields of medicine.
3.2 ๋์งํธ ํํํ (Digital Phenotyping)์ ๋ฑ์ฅ3.2 The Emergence of Digital Phenotyping
2016๋
Thomas Insel(์ NIMH ์์ฅ)์ด "digital phenotyping" ๊ฐ๋
์ ์ ์ํ ์ดํ, ์ค๋งํธํฐ์ ์๋ ์ผ์(passive sensors)์ ๋ฅ๋ ๋ฐ์ดํฐ(active data: ํ์ดํยท์์ฑ)๊ฐ ์ ์ ๊ฑด๊ฐ ์ํ์ ๊ฐ๊ด์ ์งํ๋ก ํ์ฉ ๊ฐ๋ฅํจ์ด ์
์ฆ๋์๋ค (Insel 2017; Torous et al. 2018). ๊ทธ๋ฌ๋ ๊ธฐ์กด ๋์งํธ ํํํ ์ฐ๊ตฌ๋ ๋จ์ผ ๋๋ 2-3๊ฐ ์์ค๋ง ํตํฉํ๊ณ , ๋ณธ๊ฒฉ์ ๋ค์ค ์์ค ์ตํฉ๊ณผ DSM-5-TR ์ง๋จ ์ฐ์ถ์ ๋ถ์ฌํ๋ค.
Since Thomas Insel (former NIMH director) proposed the concept of "digital phenotyping" in 2016, smartphone-passive sensors and active data (typing, voice) have been shown to serve as objective indicators of mental-health states (Insel 2017; Torous et al. 2018). However, prior digital-phenotyping research has integrated at most two or three sources, leaving full multi-source fusion and DSM-5-TR diagnostic output unaddressed.
3.3 ๋ฒ ์ด์ง์ ์์ ์์ฌ๊ฒฐ์ ์ง์์ ํ๊ณ3.3 Limits of Bayesian Clinical Decision Support
๋ฒ ์ด์ง์ ์ถ๋ก ์ ์ํ ์ง๋จ์์ 1959๋
Ledley & Lusted ์ดํ ์ ์๋ ค์ง ๋๊ตฌ์ด์ง๋ง, ์ ์ ๊ณผ์ ์ ์ฉ๋ ์ฌ๋ก๋ ๋ค์ ํ๊ณ๋ฅผ ๊ฐ์ง๋ค:
Bayesian inference has been a well-known tool in medical diagnosis since Ledley & Lusted (1959), but its applications in psychiatry have been limited as follows:
- ์ด์ง์ ๋ชจ๋ฌ๋ฆฌํฐ ์ตํฉ ๋ถ์ฌ: ์ค๋ฌธ(์ด์ฐ), qEEG(์ฐ์ยท์๊ณ์ด), ์์ฑ(์ฐ์ยท๋์ )์ ๋์์ ํตํฉํ๋ ๋ฒ ์ด์ง์ ๋ชจ๋ธ ๋ถ์ฌ.No fusion of heterogeneous modalities: no Bayesian model has integrated survey (discrete), qEEG (continuous time series), and voice (continuous dynamic) simultaneously.
- ๋จ๊ณ์ ์ ๋ขฐ๋ ์ถ์ ๋ถ์ฌ: ๋ฐ์ดํฐ๊ฐ ์ถ๊ฐ๋์ด๋ ์ ๋ขฐ๋ ๋ณํ ์ถ์ ์ด ์๋ค.No stepwise confidence tracking: no system tracks how confidence changes as data is added.
- DSM-5-TR 18 ์ง๋จ ๋์ ์ฐ์ถ ๋ถ์ฌ: ๋จ์ผ ์ง๋จ(์: ์ฐ์ธ์ฆ vs ์ ์)์ ๊ทธ์น๊ณ ๋ค์ค ์ง๋จ์ ๋์์ ํ๊ฐํ์ง ๋ชปํ๋ค.No simultaneous output of all 18 DSM-5-TR diagnoses: existing systems stop at a single diagnosis (e.g., depression vs. normal) and cannot evaluate multiple diagnoses simultaneously.
- ์์ ์ํฌํ๋ก์ฐ ํตํฉ ๋ถ์ฌ: ์๊ณ ๋ฆฌ์ฆ ๊ฒฐ๊ณผ๊ฐ ์์๊ฐ์ ๋ค์ ๊ฒฐ์ (์ถ๊ฐ ๊ฒ์ฌ ๊ถ์ , ์น๋ฃ ๊ณํ)์ผ๋ก ์ฐ๊ฒฐ๋์ง ๋ชปํ๋ค.No integration into clinical workflow: algorithmic outputs do not feed into the clinician's next decision (recommending further tests, treatment planning).
06์์ธ ์ค๋ช
Detailed Description
6.1 8 ๋ฐ์ดํฐ ์์ค์ ์ ์6.1 Definition of the Eight Data Sources
| Source |
๋ฐ์ดํฐ ์ ํData Type |
์ธก์ ๋ฐฉ๋ฒMeasurement Method |
์ฃผ์ ํน์งKey Features |
| S1: Survey | ์ด์ฐ (90 ํญ๋ชฉ)discrete (90 items) | ์๊ธฐ๋ณด๊ณ 0-4 Likertself-report Likert 0โ4 | PHQ-9, GAD-7, ASRS, PCL-5 ํตํฉintegration of PHQ-9, GAD-7, ASRS, PCL-5 |
| S2: Typing | ์๊ณ์ดtime series | ํค์คํธ๋กํฌ ํ์์คํฌํkeystroke timestamps | ์๋, ์ผ๊ด์ฑ, ์ค๋ฅ์จ, ์ ์ ๋น๋speed, consistency, error rate, correction frequency |
| S3: Phone Passive | ์๊ณ์ดtime series | ์ค๋งํธํฐ ์ผ์smartphone sensors | ๊ฐ์๋๊ณ, GPS, screen-on, ํตํ ํจํดaccelerometer, GPS, screen-on time, call patterns |
| S4: qEEG | 19์ฑ EEG19-channel EEG | EC 5min + EO 5mineyes-closed 5 min + eyes-open 5 min | TBR, FAA, PAP, COH, PAFTBR, FAA, PAP, COH, PAF |
| S5: HRV | ์ฌ๋ฐ ์๊ณ์ดheart-rate time series | 5๋ถ ์์ ์ํ5-min resting state | RMSSD, SDNN, LF/HF, pNN50 |
| S6: Face | ๋น๋์ค 30fpsvideo 30 fps | ์ผ๊ตด ๋์ ๋จ์facial action units | 17 ํต์ฌ AU ๊ฐ๋ (FACS)17 core AU intensities (FACS) |
| S7: Voice | ์ค๋์ค 44.1kHzaudio 44.1 kHz | ์์ ๋ฐํ 3๋ถ3 minutes of free speech | F0, jitter, shimmer, energy, prosody |
| S8: ERP | ์ฌ๊ฑด๊ด๋ จ์ ์event-related potentials | ์ฒญ๊ฐ/์๊ฐ oddballauditory/visual oddball | P300, N200, MMN, ERN |
6.2 18 DSM-5-TR ์ง๋จ ํด๋์ค6.2 The 18 DSM-5-TR Diagnostic Classes
์ถ๋ ฅ D = {d_1, d_2, ..., d_18}, ๊ฐ d_i๋ ๋ค์ 18๊ฐ ์ง๋จ ์ค ํ๋:
The output is D = {d_1, d_2, ..., d_18}, where each d_i is one of the following eighteen diagnoses:
- d1: ADHD โ Inattentive
- d2: ADHD โ Hyperactive/Impulsive
- d3: ADHD โ Combined
- d4: Major Depressive Disorder
- d5: Persistent Depressive Disorder
- d6: Generalized Anxiety Disorder
- d7: Panic Disorder
- d8: Social Anxiety Disorder
- d9: PTSD
- d10: OCD
- d11: Bipolar I
- d12: Bipolar II
- d13: Cyclothymic Disorder
- d14: Insomnia Disorder
- d15: Substance Use Disorder
- d16: Autism Spectrum Disorder
- d17: ํ์ต ์ฅ์ Specific Learning Disorder
- d18: ์ง๋จ ๋ฏธํด๋น (์ ์)No Diagnosis (Normal)
6.3 ๋ฒ ์ด์ง์ ๋ค์ค ๋ชจ๋ ์ตํฉ (Bayesian Multi-Module Fusion)6.3 Bayesian Multi-Module Fusion
๊ฐ ๋ชจ๋ m์ด ์ง๋จ d_i์ ๋ํด ์ฐ์ถํ๋ ์ฐ๋ L_m(d_i | data_m)๋ฅผ ๊ฒฐํฉํ๋ค:The likelihood L_m(d_i | data_m) produced by each module m for diagnosis d_i is combined as follows:
์์ 1: ๋ฒ ์ด์ง์ ์ฌํ ํ๋ฅ Equation 1: Bayesian Posterior Probability
P(d_i | D_1, D_2, ..., D_k) โ P(d_i) ยท โ_{m=1}^{k} L_m(d_i | data_m)^(w_m)
where:
P(d_i) = prior probability for diagnosis d_i
(from epidemiology or population base rates)
L_m(...) = likelihood from module m (m โ {1..8})
w_m = module reliability weight (validated on Boston Neuromind dataset)
k = number of available modules (varies by Tier)
normalize: P(d_i | data) = numerator / ฮฃ_j numerator_j for i = 1..18
6.4 4-Tier Progressive ์ ๋ขฐ๋ ์ฐ์ถ6.4 4-Tier Progressive Confidence
| Tier |
ํ์ฑ ๋ชจ๋Active Modules |
์ ๋ขฐ๋ ๋ฒ์Confidence Band |
์์ง ์๊ฐCollection Time |
ํ์ฉUse Case |
| Tier 1 โ Remote | S1 + S2 + S3 | 60โ75% | 15-30๋ถ15โ30 min | ์๊ฒฉ ์คํฌ๋ฆฌ๋remote screening |
| Tier 2 โ Clinical | + S4 (qEEG) | 75โ88% | + 15๋ถ+ 15 min | ์์ ํ์ , ๋ณดํclinical confirmation, insurance |
| Tier 3 โ Comprehensive | + S5, S6, S7 | 88โ95% | + 20๋ถ+ 20 min | ์น๋ฃ ๊ณํ, ์ถ์ treatment planning, follow-up |
| Tier 4 โ Research | + S8 + longitudinal | 95โ98% | + 30๋ถ + ์ข
๋จ+ 30 min + longitudinal | FDA 510(k), ์ถํFDA 510(k), publication |
6.5 ์ ๋ขฐ๋ ๋จ์กฐ ์ฆ๊ฐ ๋ณด์ฆ6.5 Monotonic Confidence Guarantee
์๊ณ ๋ฆฌ์ฆ์ ๋ค์์ ์ํ์ ์ผ๋ก ๋ณด์ฆํ๋ค:The algorithm mathematically guarantees the following:
์์ 2: ๋จ์กฐ ์ฆ๊ฐ ์ ๋ฆฌEquation 2: Monotonicity Theorem
H(D | Tier_n+1) โค H(D | Tier_n)
where:
H(D | Tier_n) = Shannon entropy of diagnostic distribution at Tier n
์ฆ, Tier๊ฐ ์์นํ ๋ ์ง๋จ ๋ถํฌ์ ๋ถํ์ค์ฑ(์ํธ๋กํผ)์ ๊ฒฐ์ฝ ์ฆ๊ฐํ์ง ์๋๋ค.
This means uncertainty (entropy) of the diagnostic distribution
never increases as Tier rises.
Confidence(Tier_n) = 1 - H(D|Tier_n) / H_max, where H_max = log_2(18)
์ด ๋จ์กฐ์ฑ ๋ณด์ฆ์ด ๋ณธ ๋ฐ๋ช
์ ํต์ฌ์ด๋ค โ ์์๊ฐ์๊ฒ "๋ฐ์ดํฐ๋ฅผ ๋ ์์งํ ์๋ก ์ง๋จ ์ ๋ขฐ๋๊ฐ ๊ฒฐ์ฝ ๋จ์ด์ง์ง ์๋๋ค"๋ ์ํ์ ๋ณด์ฅ์ ์ ๊ณตํ๋ค.This monotonicity guarantee is at the core of the invention: it provides clinicians the mathematical assurance that "diagnostic confidence will never decrease as more data is collected."
6.6 Tier ์
๊ทธ๋ ์ด๋ ์ถ์ฒ๊ธฐ (TUR)6.6 Tier Upgrade Recommender (TUR)
ํ์ฌ Tier์์ ๋ค์ Tier๋ก ์
๊ทธ๋ ์ด๋ ์ ์์ ์ ๋ขฐ๋ ์์น์ ์ฌ์ ์ฐ์ถํ๋ค:The TUR pre-computes the expected confidence gain from upgrading the current Tier to the next:
ProjectedGain(Tier_n โ Tier_n+1) =
E[Confidence(Tier_n+1)] - Confidence(Tier_n)
โ "Adding qEEG (Tier 2) is projected to raise confidence
from 68% to 81% (+13pp). Recommended."
6.7 ์์ ์ํฌํ๋ก์ฐ ํตํฉ6.7 Clinical Workflow Integration
- ํ์ ๋ฑ๋ก โ Tier 1 ๋ฐ์ดํฐ ์์ง (์ค๋ฌธ, ํ์ดํ, ํฐ)Patient enrollment โ Tier 1 data collection (survey, typing, phone)
- BMF ์์ง ์คํ โ 18 ์ง๋จ๋ณ ์ฌํ ํ๋ฅ ์ฐ์ถ โ Top 3 ์ง๋จ + ์ ๋ขฐ๋Run BMF engine โ compute posterior probabilities for the 18 diagnoses โ return top 3 with confidence
- TUR ๊ถ๊ณ โ "qEEG ์ถ๊ฐ ์ ์ ๋ขฐ๋ +X% ์์น" ํ์TUR recommendation โ display "adding qEEG raises confidence by +X %"
- ์์๊ฐ ๊ฒฐ์ โ Tier 2/3/4๋ก ์
๊ทธ๋ ์ด๋ ๋๋ ํ Tier์์ ์ง๋จ ํ์ Clinician decision โ upgrade to Tier 2/3/4 or finalize diagnosis at the current Tier
- ์ต์ข
๋ณด๊ณ ์ โ ์ง๋จ + ์ ๋ขฐ๋ + ๋ฐ์ดํฐ ์์ค๋ณ ๊ธฐ์ฌ๋(transparency)Final report โ diagnosis + confidence + per-source contribution (transparency)
08์ฒญ๊ตฌํญClaims
์ฒญ๊ตฌํญ 1 (๋
๋ฆฝํญ)Claim 1 (Independent)
ํ์์ ๋ํ DSM-5-TR ์ ์ ๊ณผ ์ง๋จ์ ๋ค์ค ๋ฐ์ดํฐ ์์ค๋ก๋ถํฐ ์ฐ์ถํ๋, ์ปดํจํฐ ๊ตฌํ ๋ฐฉ๋ฒ์ผ๋ก์:
(a) ํ์๋ก๋ถํฐ ๋ณต์์ ์ด์ง์ ๋ฐ์ดํฐ ์์ค โ ์๊ธฐ๋ณด๊ณ ์ค๋ฌธ, ํ์ดํ ํ๋ ๋ฐ์ดํฐ, ์ค๋งํธํฐ ์๋ ์ผ์ ๋ฐ์ดํฐ, 19์ฑ๋ ์ ๋๋ํ(qEEG), ์ฌ๋ฐ๋ณ์ด๋(HRV), ์ผ๊ตด ๋์ ๋จ์(FAU), ์์ฑ ์ด์จ ๋ฐ ์ฌ๊ฑด๊ด๋ จ์ ์(ERP) ์ค ์ ์ด๋ 4๊ฐ โ ๋ฅผ ์์ ํ๋ ๋จ๊ณ;
(b) ๊ฐ ๋ฐ์ดํฐ ์์ค์ ๋ํด ๋ณ๊ฐ์ ๋ชจ๋์ ์ ์ฉํ์ฌ, 18๊ฐ DSM-5-TR ์ง๋จ ํด๋์ค ๊ฐ๊ฐ์ ๋ํ ์ฐ๋(likelihood)๋ฅผ ์ฐ์ถํ๋ ๋จ๊ณ;
(c) ๊ฐ ๋ชจ๋๋ก๋ถํฐ ์ฐ์ถ๋ ์ฐ๋๋ฅผ ๋ฒ ์ด์ง์ ๋ค์ค ๋ชจ๋ ์ตํฉ ์์ง์ ์
๋ ฅํ์ฌ, ์ฌ์ ํ๋ฅ ๊ณผ ๋ชจ๋๋ณ ์ ๋ขฐ๋ ๊ฐ์ค์น์ ๊ธฐ๋ฐํ ์ฌํ ํ๋ฅ ์ 18๊ฐ ์ง๋จ ํด๋์ค ๊ฐ๊ฐ์ ๋ํด ์ฐ์ถํ๋ ๋จ๊ณ;
(d) ์ฐ์ถ๋ 18๊ฐ ์ฌํ ํ๋ฅ ๋ถํฌ์ ์๋ ์ํธ๋กํผ๋ก๋ถํฐ ์ง๋จ ์ ๋ขฐ๋๋ฅผ ์ฐ์ถํ๋, ํด๋น ์ ๋ขฐ๋๋ ํ์ฑํ๋ ๋ฐ์ดํฐ ์์ค์ ์์ ๋ํ ํจ์๋ก์ 4๋จ๊ณ์ ๊ณ์ธต์ ์ ๋ขฐ๋ ๋ฒ์ โ Tier 1: 60% ์ด์ 75% ์ดํ, Tier 2: 75% ์ด์ 88% ์ดํ, Tier 3: 88% ์ด์ 95% ์ดํ, Tier 4: 95% ์ด์ 98% ์ดํ โ ์ค ํ๋์ ์ํ๋ ๋จ๊ณ;
(e) ์ถ๊ฐ ๋ฐ์ดํฐ ์์ค ํตํฉ ์ ์์๋๋ ์ ๋ขฐ๋ ์์น๊ฐ์ ์ฌ์ ์ฐ์ถํ๊ณ , ์์๊ฐ์๊ฒ ๋ค์ Tier๋ก์ ์
๊ทธ๋ ์ด๋ ๊ถ๊ณ ๋ฅผ ์ถ๋ ฅํ๋ ๋จ๊ณ; ๋ฐ
(f) ๊ฐ์ฅ ๋์ ์ฌํ ํ๋ฅ ์ ๊ฐ์ง๋ ์ง๋จ ๋ฐ ๊ทธ ์ ๋ขฐ๋๋ฅผ ์ถ๋ ฅํ๋ ๋จ๊ณ;
๋ฅผ ํฌํจํ๋ ๋ฐฉ๋ฒ.
A computer-implemented method for producing a DSM-5-TR psychiatric diagnosis for a patient from multiple data sources, the method comprising:
(a) receiving from the patient at least four of a plurality of heterogeneous data sources comprising self-report survey, typing-behavior data, smartphone-passive sensor data, 19-channel quantitative EEG (qEEG), heart-rate variability (HRV), Facial Action Units (FAU), vocal prosody, and event-related potentials (ERP);
(b) applying a distinct module to each data source to compute likelihoods for each of eighteen DSM-5-TR diagnostic classes;
(c) inputting the likelihoods computed by each module into a Bayesian Multi-Module Fusion engine, the engine computing posterior probabilities for each of the eighteen diagnostic classes based on prior probabilities and module-specific reliability weights;
(d) computing diagnostic confidence from the Shannon entropy of the resulting eighteen-class posterior distribution, said confidence falling within one of four hierarchical confidence bands as a function of the number of active data sources โ Tier 1: 60 % to 75 %; Tier 2: 75 % to 88 %; Tier 3: 88 % to 95 %; Tier 4: 95 % to 98 %;
(e) pre-computing the projected confidence gain from incorporating an additional data source, and outputting to a clinician a recommendation to upgrade to the next Tier; and
(f) outputting the diagnosis with the highest posterior probability together with its confidence value.
์ฒญ๊ตฌํญ 2 (์ข
์ํญ)Claim 2 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (c)์ ์ฌํ ํ๋ฅ ์ ๋ค์ ์์ ์ํด ์ฐ์ถ๋๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ: P(d_i | data) โ P(d_i) ยท โ_{m=1}^{k} L_m(d_i | data_m)^(w_m), ์ฌ๊ธฐ์ P(d_i)๋ ์ง๋จ d_i์ ์ฌ์ ํ๋ฅ , L_m์ ๋ชจ๋ m์ ์ฐ๋, w_m์ ๋ชจ๋ m์ ์ ๋ขฐ๋ ๊ฐ์ค์น, k๋ ํ์ฑ ๋ชจ๋ ์์. ๋ํ 18๊ฐ ์ง๋จ ํด๋์ค์ ๋ํด ์ ๊ทํ๋๋ค.
The method of Claim 1, wherein the posterior probability of step (c) is computed by P(d_i | data) โ P(d_i) ยท โ_{m=1}^{k} L_m(d_i | data_m)^(w_m), where P(d_i) is the prior probability of diagnosis d_i, L_m is the likelihood from module m, w_m is the reliability weight of module m, and k is the number of active modules; the result is normalized over the eighteen diagnostic classes.
์ฒญ๊ตฌํญ 3 (์ข
์ํญ)Claim 3 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (d)์ ์ง๋จ ์ ๋ขฐ๋๋ ๋ฐ์ดํฐ ์์ค๊ฐ ์ถ๊ฐ๋ ๋ ๋จ์กฐ ๋น์ฆ๊ฐ์ฑ(monotonic non-increase)์ ๋ง์กฑํ๋ ์ฌํ ๋ถํฌ์ ์๋ ์ํธ๋กํผ H(D)์ ๊ธฐ๋ฐํ๋ฉฐ, ์ฆ H(D | Tier_n+1) โค H(D | Tier_n) ์ ์ํ์ ๋ณด์ฅ์ ๋ฐ๋ฅด๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein the diagnostic confidence of step (d) is based on the Shannon entropy H(D) of the posterior distribution that satisfies a monotonic-non-increase property as data sources are added, namely the mathematical guarantee H(D | Tier_n+1) โค H(D | Tier_n).
์ฒญ๊ตฌํญ 4 (์ข
์ํญ)Claim 4 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (a)์ 18๊ฐ DSM-5-TR ์ง๋จ ํด๋์ค๋ ADHD ๋ถ์ฃผ์ํ, ADHD ๊ณผ์ํ๋/์ถฉ๋ํ, ADHD ๋ณตํฉํ, ์ฃผ์์ฐ์ธ์ฅ์ , ์ง์์ฑ์ฐ์ธ์ฅ์ , ๋ฒ๋ถ์์ฅ์ , ๊ณตํฉ์ฅ์ , ์ฌํ๋ถ์์ฅ์ , ์ธ์ํ์คํธ๋ ์ค์ฅ์ , ๊ฐ๋ฐ์ฅ์ , ์๊ทน์ฑI์ฅ์ , ์๊ทน์ฑII์ฅ์ , ์ํ์ฑ์ฅ์ , ๋ถ๋ฉด์ฅ์ , ๋ฌผ์ง์ฌ์ฉ์ฅ์ , ์ํ์คํํธ๋ผ์ฅ์ , ํน์ ํ์ต์ฅ์ ๋ฐ ์ง๋จ ๋ฏธํด๋น์ ํฌํจํ๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein the eighteen DSM-5-TR diagnostic classes of step (a) comprise ADHD-Inattentive, ADHD-Hyperactive/Impulsive, ADHD-Combined, Major Depressive Disorder, Persistent Depressive Disorder, Generalized Anxiety Disorder, Panic Disorder, Social Anxiety Disorder, Posttraumatic Stress Disorder, Obsessive-Compulsive Disorder, Bipolar I Disorder, Bipolar II Disorder, Cyclothymic Disorder, Insomnia Disorder, Substance Use Disorder, Autism Spectrum Disorder, Specific Learning Disorder, and No-Diagnosis.
์ฒญ๊ตฌํญ 5 (์ข
์ํญ)Claim 5 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, Tier 1์ ์๊ธฐ๋ณด๊ณ ์ค๋ฌธ, ํ์ดํ ๋ฐ์ดํฐ, ์ค๋งํธํฐ ์๋ ๋ฐ์ดํฐ๋ฅผ ํฌํจํ๊ณ ; Tier 2๋ ์ถ๊ฐ๋ก 19์ฑ๋ qEEG๋ฅผ ํฌํจํ๊ณ ; Tier 3์ ์ถ๊ฐ๋ก HRV, ์ผ๊ตด ๋์ ๋จ์ ๋ฐ ์์ฑ ์ด์จ์ ํฌํจํ๊ณ ; Tier 4๋ ์ถ๊ฐ๋ก ์ฌ๊ฑด๊ด๋ จ์ ์(ERP) ๋ฐ ์ข
๋จ์ ์ธก์ ์ ํฌํจํ๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein Tier 1 comprises self-report survey, typing data, and smartphone-passive data; Tier 2 additionally comprises 19-channel qEEG; Tier 3 additionally comprises HRV, Facial Action Units, and vocal prosody; and Tier 4 additionally comprises event-related potentials (ERP) and longitudinal measurements.
์ฒญ๊ตฌํญ 6 (์ข
์ํญ)Claim 6 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (e)์ Tier ์
๊ทธ๋ ์ด๋ ๊ถ๊ณ ๋ ProjectedGain(Tier_n โ Tier_n+1) = E[Confidence(Tier_n+1)] โ Confidence(Tier_n)์ ์ ๋์ ์ ๋ขฐ๋ ์์น๊ฐ์ ์ฐ์ถํ๊ณ , ํด๋น ์์น๊ฐ์ด ๋ฏธ๋ฆฌ ์ ํด์ง ์๊ณ์น(์: 5 ๋ฐฑ๋ถ์จ ํฌ์ธํธ)๋ฅผ ์ด๊ณผํ๋ ๊ฒฝ์ฐ์๋ง ์์๊ฐ์๊ฒ ๊ถ๊ณ ๋ฅผ ์ถ๋ ฅํ๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein the Tier-upgrade recommendation of step (e) computes the quantitative confidence gain ProjectedGain(Tier_n โ Tier_n+1) = E[Confidence(Tier_n+1)] โ Confidence(Tier_n), and outputs a recommendation to the clinician only when said gain exceeds a predefined threshold (for example, five percentage points).
์ฒญ๊ตฌํญ 7 (์ข
์ํญ)Claim 7 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (f)์ ์ถ๋ ฅ์ ๋จ์ผ ์ง๋จ๋ฟ๋ง ์๋๋ผ, ์ฌํ ํ๋ฅ ์๊ณ์น(์: 30%)๋ฅผ ์ด๊ณผํ๋ ๋ชจ๋ ์ง๋จ์ ๊ณต์กด(comorbid) ์ง๋จ ํ๋ณด๋ก ํจ๊ป ์ถ๋ ฅํ๊ณ , ๊ฐ ์ง๋จ๋ณ๋ก ์ด๋ค ๋ฐ์ดํฐ ์์ค๊ฐ ์ด๋ ์ ๋ ๊ธฐ์ฌํ๋์ง์ ๋ฐ์ดํฐ ์์ค๋ณ ๊ธฐ์ฌ๋(transparency report)๋ฅผ ํฌํจํ๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein the output of step (f) includes not only a single diagnosis but also any diagnosis exceeding a posterior-probability threshold (for example, 30 %) as a candidate comorbid diagnosis, and includes a per-source contribution report (transparency report) indicating how much each data source contributed to each diagnosis.
์ฒญ๊ตฌํญ 8 (์ข
์ํญ)Claim 8 (Dependent)
์ฒญ๊ตฌํญ 1์ ์์ด์, ๋จ๊ณ (b)์ ๋ชจ๋๋ณ ์ฐ๋ ์ฐ์ถ์ ์ฌ์ ํ์ต๋ ๋จธ์ ๋ฌ๋ ๋ชจ๋ธ, ์ ์ ๋ชจ์ง๋จ ๋๋น Z-์ ์, ๋๋ ์์ ๊ฒ์ฆ๋ ๋ฃฐ ๊ธฐ๋ฐ ๋งคํ ์ค ํ๋ ๋๋ ๊ทธ ์ด์์ ์กฐํฉ์ผ๋ก ์ํ๋๋ ๊ฒ์ ํน์ง์ผ๋ก ํ๋ ๋ฐฉ๋ฒ.
The method of Claim 1, wherein the per-module likelihood computation of step (b) is performed by one or a combination of pre-trained machine-learning models, Z-scoring against a normative population, and clinically validated rule-based mappings.
์ฒญ๊ตฌํญ 9 (๋
๋ฆฝํญ โ ์์คํ
)Claim 9 (Independent โ System)
์ฒญ๊ตฌํญ 1 ๋ด์ง 8 ์ค ์ด๋ ํ ํญ์ ๋ฐฉ๋ฒ์ ์ํํ๊ธฐ ์ํ, ์ ์ด๋ ํ๋์ ํ๋ก์ธ์, 8๊ฐ ์ด์ง์ ๋ฐ์ดํฐ ์์ค๋ฅผ ์์ ํ๋ ๋ค์ค ๋ชจ๋ฌ ์
๋ ฅ ์ธํฐํ์ด์ค, ๋ฐ ์๊ธฐ ํ๋ก์ธ์์ ์ํด ์คํ๋๋ ๋ช
๋ น์ด๋ฅผ ์ ์ฅํ๋ ๋น์ผ์์ ์ปดํจํฐ ํ๋
๊ฐ๋ฅ ์ ์ฅ ๋งค์ฒด๋ฅผ ํฌํจํ๋ 4-Tier Progressive ๋ค์ค ์์ค ์ง๋จ ์์คํ
.
A 4-Tier Progressive multi-source diagnostic system for performing the method of any one of Claims 1 through 8, the system comprising at least one processor, a multi-modal input interface for receiving the eight heterogeneous data sources, and a non-transitory computer-readable storage medium storing instructions executable by the processor.
์ฒญ๊ตฌํญ 10 (๋
๋ฆฝํญ โ ๋งค์ฒด)Claim 10 (Independent โ Medium)
์ปดํจํฐ์ ์ํด ์คํ๋ ๋ ์ฒญ๊ตฌํญ 1 ๋ด์ง 8 ์ค ์ด๋ ํ ํญ์ ๋ฐฉ๋ฒ์ ์ํํ๋๋ก ํ๋ ๋ช
๋ น์ด๋ฅผ ์ ์ฅํ๋ ๋น์ผ์์ ์ปดํจํฐ ํ๋
๊ฐ๋ฅ ์ ์ฅ ๋งค์ฒด.
A non-transitory computer-readable storage medium storing instructions which, when executed by a computer, cause the computer to perform the method of any one of Claims 1 through 8.
11๊ด๋ จ ๋
ผ๋ฌธReferences
๋ณธ ๋ฐ๋ช
์ ์ด๋ก ์ ยท์์์ ๊ทผ๊ฑฐ๊ฐ ๋๋ ํต์ฌ ๋
ผ๋ฌธ ๋ฐ ์๋ฃ. ํด๋ฆญํ๋ฉด ์ธ๋ถ ์ถ์ฒ๋ก ์ด๋ํฉ๋๋ค.
Key papers and resources providing the theoretical and clinical basis for this invention. Click links to access external sources.
A. DSM-5-TR ๋ฐ RDoC ํ๋ ์์ํฌA. DSM-5-TR and RDoC Framework
- A1American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR). Washington, DC: APA, 2022. DSM-5-TR โ
- A2Insel TR. The NIMH Research Domain Criteria (RDoC) Project: Precision Medicine for Psychiatry. American Journal of Psychiatry, 2014; 171(4):395-397. DOI โ
- A3Cuthbert BN, Insel TR. Toward the future of psychiatric diagnosis: the seven pillars of RDoC. BMC Medicine, 2013; 11:126. DOI โ
- A4Lieblich SM, Castle DJ, Pantelis C, et al. High heterogeneity and low reliability in the diagnosis of major depression. BJPsych Open, 2015; 1(2):e5-e7. DOI โ
B. ๋ฒ ์ด์ง์ ์ง๋จ ์ถ๋ก B. Bayesian Diagnostic Inference
- B1Ledley RS, Lusted LB. Reasoning Foundations of Medical Diagnosis. Science, 1959; 130(3366):9-21. DOI โ
- B2Pearl J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, 1988. Book โ
- B3Heckerman D, Horvitz E, Nathwani BN. Toward Normative Expert Systems: Part I. The Pathfinder Project. Methods of Information in Medicine, 1992; 31(2):90-105. PubMed โ
- B4Shannon CE. A Mathematical Theory of Communication. Bell System Technical Journal, 1948; 27:379-423, 623-656. DOI โ
C. QEEG ์์ ์์ฉC. QEEG Clinical Applications
- C1Kropotov JD. Quantitative EEG, Event-Related Potentials and Neurotherapy. Academic Press, 2009. Book โ
- C2Thatcher RW. Coherence, Phase Differences, Phase Shift, and Phase Lock in EEG/ERP Analyses. Developmental Neuropsychology, 2012; 37(6):476-496. DOI โ
- C3Arns M, Conners CK, Kraemer HC. A decade of EEG Theta/Beta Ratio Research in ADHD: a meta-analysis. Journal of Attention Disorders, 2013; 17(5):374-383. DOI โ
- C4Henriques JB, Davidson RJ. Left frontal hypoactivation in depression. Journal of Abnormal Psychology, 1991; 100(4):535-545. DOI โ
- C5Hammond DC. What is Neurofeedback: An Update. Journal of Neurotherapy, 2011; 15(4):305-336. DOI โ
D. ๋์งํธ ํํํ (Digital Phenotyping)D. Digital Phenotyping
- D1Insel TR. Digital Phenotyping: Technology for a New Science of Behavior. JAMA, 2017; 318(13):1215-1216. DOI โ
- D2Torous J, Kiang MV, Lorme J, Onnela JP. New Tools for New Research in Psychiatry: A Scalable and Customizable Platform to Empower Data Driven Smartphone Research. JMIR Mental Health, 2016; 3(2):e16. DOI โ
- D3Onnela JP, Rauch SL. Harnessing Smartphone-Based Digital Phenotyping to Enhance Behavioral and Mental Health. Neuropsychopharmacology, 2016; 41:1691-1696. DOI โ
- D4Mohr DC, Zhang M, Schueller SM. Personal Sensing: Understanding Mental Health Using Ubiquitous Sensors and Machine Learning. Annual Review of Clinical Psychology, 2017; 13:23-47. DOI โ
E. HRV ๋ฐ ์์จ์ ๊ฒฝ๊ณE. HRV and Autonomic Nervous System
- E1Thayer JF, Lane RD. The role of vagal function in the risk for cardiovascular disease and mortality. Biological Psychology, 2007; 74(2):224-242. DOI โ
- E2Shaffer F, Ginsberg JP. An Overview of Heart Rate Variability Metrics and Norms. Frontiers in Public Health, 2017; 5:258. DOI โ
- E3Beauchaine TP, Thayer JF. Heart rate variability as a transdiagnostic biomarker of psychopathology. International Journal of Psychophysiology, 2015; 98(2):338-350. DOI โ
F. ์ผ๊ตดยท์์ฑ ์ ์ ๋ถ์F. Facial and Vocal Affect Analysis
- F1Ekman P, Friesen WV. Facial Action Coding System (FACS): A Technique for the Measurement of Facial Movement. Consulting Psychologists Press, 1978.
- F2Cohn JF, Kruez TS, Matthews I, et al. Detecting depression from facial actions and vocal prosody. 3rd International Conference on Affective Computing and Intelligent Interaction (ACII), 2009. DOI โ
- F3Cummins N, Scherer S, Krajewski J, et al. A review of depression and suicide risk assessment using speech analysis. Speech Communication, 2015; 71:10-49. DOI โ
- F4Mundt JC, Snyder PJ, Cannizzaro MS, et al. Voice acoustic measures of depression severity and treatment response. Journal of Neurolinguistics, 2007; 20(1):50-64. DOI โ
G. ํ์ดํ ํ๋ ํจํด (Behavioral Biometrics)G. Typing Behavior Patterns (Behavioral Biometrics)
- G1Mastoras RE, Iakovakis D, Hadjidimitriou S, et al. Touchscreen typing pattern analysis for remote detection of the depressive tendency. Scientific Reports, 2019; 9:13414. DOI โ
- G2Cao B, Zheng L, Zhang C, et al. DeepMood: Modeling Mobile Phone Typing Dynamics for Mood Detection. KDD '17 Conference Proceedings, 2017. DOI โ
H. ERP ์์ ์์ฉH. ERP Clinical Applications
- H1Polich J. Updating P300: An integrative theory of P3a and P3b. Clinical Neurophysiology, 2007; 118(10):2128-2148. DOI โ
- H2Bramon E, Rabe-Hesketh S, Sham P, et al. Meta-analysis of the P300 and P50 waveforms in schizophrenia. Schizophrenia Research, 2004; 70(2-3):315-329. DOI โ
I. FDA / ๊ท์ ๊ฐ์ด๋๋ผ์ธI. FDA / Regulatory Guidelines
- I1U.S. Food and Drug Administration. Software as a Medical Device (SaMD): Clinical Evaluation Guidance. FDA, 2017. FDA โ
- I2U.S. Food and Drug Administration. Artificial Intelligence and Machine Learning in Software as a Medical Device. FDA โ