2nd/10_Wiki/Topic_Programming/AI_and_ML/Neuroprosthetics-Development.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-neuroprosthetics-development	Neuroprosthetics Development	10_Wiki/Topics	verified	self	Neuroprosthetics BCI Brain-Computer Interface Neural Prosthesis	none	A	0.9	applied	bci neuroprosthetics neuralink synchron cochlear retinal-prosthesis motor-bci		2026-05-10	pending	language framework python pytorch-mne

Neuroprosthetics Development

매 한 줄

• 신경보철은 신경계와 직접 인터페이스하여 손실된 감각·운동·인지 기능을 복원하는 기기군이다(BCI, cochlear, retinal, motor prosthesis).

매 핵심

• 분류: (1) 감각 보철(cochlear, retinal), (2) 운동 BCI(invasive Utah array, ECoG, Stentrode), (3) 인지/심부자극(DBS for PD, OCD).
• 2026 현황: Neuralink N1 첫 인간 임플란트(2024) → 1024ch threadlike electrode + bluetooth, Synchron Stentrode(stent 형태, 정맥경유, 16ch), Blackrock Utah array(96–256ch, gold standard), Cochlear Nucleus 8(~22 ch electrode).
• 신호 처리 파이프라인: spike sorting → feature(firing rate, LFP power) → Kalman/RNN decoder → effector(cursor, robot arm, speech).
• Motor BCI breakthrough: ALS 환자 speech BCI(Stanford 2023) ~62 wpm, BrainGate 8자유도 로봇팔.
• 재료/수명: 만성 임플란트 6–12개월 후 glial encapsulation, signal degradation. 유연 polymer probe(NeuroNexus, Paradromics)로 개선.

💻 패턴

# Spike sorting with template matching (simplified)
import numpy as np
from scipy.signal import butter, filtfilt

def bandpass(x, fs=30000, lo=300, hi=6000):
    b, a = butter(4, [lo / (fs / 2), hi / (fs / 2)], btype="band")
    return filtfilt(b, a, x)

def detect_spikes(x, thresh_sd=4):
    thr = -thresh_sd * np.median(np.abs(x)) / 0.6745
    return np.where(x < thr)[0]

# Kalman filter decoder: neural firing → cursor velocity
import numpy as np

class KalmanDecoder:
    def __init__(self, A, C, W, Q):
        self.A, self.C, self.W, self.Q = A, C, W, Q
        self.x = np.zeros(A.shape[0])
        self.P = np.eye(A.shape[0])

    def step(self, y):
        # predict
        self.x = self.A @ self.x
        self.P = self.A @ self.P @ self.A.T + self.W
        # update
        K = self.P @ self.C.T @ np.linalg.inv(self.C @ self.P @ self.C.T + self.Q)
        self.x = self.x + K @ (y - self.C @ self.x)
        self.P = (np.eye(len(self.x)) - K @ self.C) @ self.P
        return self.x  # [vx, vy]

# RNN decoder for speech BCI (handwriting/speech-to-text from cortex)
import torch, torch.nn as nn

class CortexRNN(nn.Module):
    def __init__(self, n_channels=256, hidden=512, n_phonemes=39):
        super().__init__()
        self.rnn = nn.GRU(n_channels, hidden, num_layers=2, batch_first=True)
        self.head = nn.Linear(hidden, n_phonemes)

    def forward(self, x):  # (B, T, C)
        h, _ = self.rnn(x)
        return self.head(h)  # CTC loss downstream

# Cochlear implant: CIS strategy (continuous interleaved sampling)
import numpy as np
from scipy.signal import hilbert

def cis_encode(audio, n_channels=22, fs=16000):
    bands = np.linspace(200, 8000, n_channels + 1)
    pulses = []
    for i in range(n_channels):
        # bandpass + envelope (Hilbert)
        from scipy.signal import butter, filtfilt
        b, a = butter(4, [bands[i] / (fs / 2), bands[i + 1] / (fs / 2)], btype="band")
        env = np.abs(hilbert(filtfilt(b, a, audio)))
        pulses.append(env)
    return np.array(pulses)  # delivered as biphasic pulse trains

# Retinal prosthesis (Argus II-style): downsample + polarity coding
import numpy as np

def retinal_encode(image_gray, n_electrodes=60):
    h, w = image_gray.shape
    grid = int(np.sqrt(n_electrodes))
    block_h, block_w = h // grid, w // grid
    out = np.zeros((grid, grid))
    for i in range(grid):
        for j in range(grid):
            out[i, j] = image_gray[i*block_h:(i+1)*block_h, j*block_w:(j+1)*block_w].mean()
    return out  # → electrode current amplitude

# Closed-loop DBS: detect beta burst (PD) and trigger stimulation
def beta_burst_trigger(lfp, fs=1000, lo=13, hi=30, thresh=2.0):
    from scipy.signal import butter, filtfilt
    import numpy as np
    b, a = butter(4, [lo / (fs / 2), hi / (fs / 2)], btype="band")
    beta = filtfilt(b, a, lfp)
    env = np.abs(beta)
    return env > thresh * env.std()  # boolean per sample

# Online recalibration: ridge regression refit every block
from sklearn.linear_model import Ridge
def recalibrate(X_block, y_block, alpha=1.0):
    return Ridge(alpha=alpha).fit(X_block, y_block)

매 결정 기준

• 침습 vs 비침습: 고대역폭(speech, robot arm) → invasive(Utah, Neuralink). 보조 통신·간단 cursor → ECoG/Stentrode.
• Decoder: 저차원 cursor → Kalman. 고차원 sequence(speech, handwriting) → RNN/Transformer + CTC.
• 재료: 단기 임상 → silicon Utah. 만성·유연성 → polyimide/PEDOT:PSS.
• 윤리/규제: FDA IDE, IRB, informed consent. 결과 발표 전 explanted device 분석 필수.

🔗 Graph

• 관련: Neurorehabilitation-Post-Stroke, Neuroplasticity

🤖 LLM 활용

• 임상 protocol 초안 검토(IRB 양식 비교).
• 임플란트 후 환자 보고 데이터 요약.
• decoder hyperparameter 탐색 제안(grid spec).

❌ 안티패턴

• 비정형 spike sorting을 임상 결정에 직접 사용.
• 만성 임플란트 noise drift 보정 없는 고정 decoder.
• 환자 home use에서 fail-safe(자극 정지 버튼) 부재.

🧪 검증

• BCI bench: cursor BPS(bits-per-second), word error rate(speech BCI).
• 안전: impedance trend, infection rate, MRI compatibility.

🕓 Changelog

• 2026-05-08 Phase 1: 초안 자동 생성.
• 2026-05-10 Manual cleanup: 본문 보강, Neuralink/Synchron 2026 현황 반영, 코드 패턴 7개.