Subject-Wise Depression Screening from Eight-Channel Resting-State EEG Using Asymmetry-Aware Spectral Features and Connectivity Ablation.

Code ↔ Paper

10 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 10 matches

[1] § 2. Materials and Methods › 2.4. Classifiers ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 346–401 · score 0.97 · depthwise spatial convolution, temporal kernel length, pointwise filters, separable convolution, temporal filters, depth multiplier
[2] § 2. Materials and Methods › 2.4. Classifiers ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 628–737 · score 0.78 · MLPClassifier, hidden layers, validation fraction, Extra Trees, activations, baseline
[3] § 2. Materials and Methods › 2.2. Channel Selection and Signal Processing ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 447–530 · score 0.77 · inner training split, training partition, score normalisation, epochs, validation, fitted
[4] § 2. Materials and Methods › 2.5. Evaluation Protocol and Experimental Design ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 875–987 · score 0.71 · fitted Extra Trees, Gini importance, importance ranking, evaluated Extra Trees, dimensional spectral, probed
[5] § 3. Results › 3.3. Spectral Feature-Selection Ablation ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 875–987 · score 0.68 · Gini importance ranking, feature selection ablation, random seed, Extra Trees, row, refitted
[6] § 2. Materials and Methods › 2.4. Classifiers ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 312–343 · score 0.66 · ReLU, adaptive, max, kernel, head, linear
[7] § 3. Results › 3.1. Compact Eight-Channel Baselines ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 1098–1155 · score 0.66 · raw signal deep, EEGNet, ShallowConvNet, balanced accuracy, Extra Trees, spectral feature
[8] § 2. Materials and Methods › 2.3. Feature Extraction › 2.3.2. Connectivity Features and Fusion ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 116–171 · score 0.58 · 12–30 Hz, 8–12 Hz, beta, bands, alpha, connectivity
[9] § 2. Materials and Methods › 2.5. Evaluation Protocol and Experimental Design ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 537–553 · score 0.57 · percentile bootstrap, confidence interval, bounds, resamples
[10] § 2. Materials and Methods › 2.3. Feature Extraction › 2.3.1. Spectral Features ↔ EEG_Depression_Analysis_Public_Notebook.ipynb, lines 116–171 · score 0.56 · 12–30 Hz, 8–12 Hz, beta, bands, alpha, channel

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The paper is loaded by your browser, with JavaScript: read it at doi.org or on Europe PMC.

The authors' code

Jupyter notebook · 1,297 lines · 48 KB · no license · 10 matches

# %% [markdown]
# # EEG-Based Depression Screening from 8-Channel Resting-State EEG
# by Hassan Ugail, Newton Howard, Ali Ahmed Elmahmudi and Zied Mnasri
#
# ## Subject-wise evaluation, compact baselines, connectivity ablation, deep-learning baselines
#
# Cite as: H Ugail, N Howard, A A Elmahmudi and Z Mnasri, Subject-Wise Depression Screening from 8-Channel Resting-State EEG using Asymmetry-Aware Spectral Features and Connectivity Ablation, Sensors, To Appear, 2026.
#
# ## What this notebook does
# It loads the processed outputs from the preprocessing notebook and reproduces the final experimental package:
#
# 1. **Correct subject identity handling**
# - uses `subject_key = label + "_" + subject_id`
# - avoids merging `H_S1` and `MDD_S1`
#
# 2. **Experiment 1 — Compact 8-channel baselines**
# - Extra Trees on asymmetry-aware spectral features
# - MLP on the same feature vector
# - compact 1D CNN on raw 8-channel segments
# - **EEGNet** (Lawhern et al., 2018) on raw 8-channel segments
# - **ShallowConvNet** (Schirrmeister et al., 2017) on raw 8-channel segments
#
# 3. **Experiment 2 — Connectivity ablation**
# - Extra Trees on spectral features only
# - Extra Trees on connectivity features only
# - Extra Trees on spectral + connectivity fusion
# - MLP on spectral + connectivity fusion
#
# 4. **Experiment 3 — Spectral feature-selection ablation**
# - Top-$K$ Extra Trees over $K \in \{5, 10, 15, 20, 30, 50, 70, 90\}$ features
# - Tests whether the 90-dimensional spectral vector is over-parameterised
#
# 5. **Bootstrap inference**
# - 95% percentile bootstrap confidence intervals of the mean across the 10 repeats
# - Reported for every metric in every experiment
#
# 6. **Publication-ready outputs**
# - manuscript tables with both SD and 95% CI columns
# - compact results CSV files
# - high-resolution figures with non-overlapping labels and 95% CI error bars
#
# ## Dataset
# This notebook assumes you already generated the processed files from the public dataset:
#
# **Mumtaz, Wajid (2016). _MDD Patients and Healthy Controls EEG Data (New)._ figshare. Dataset.**
# DOI: `10.6084/m9.figshare.4244171.v2`
#
# ## Input assumption
# This notebook expects the processed outputs from the earlier preprocessing stage under:
#
# `/content/drive/MyDrive/.../run_subjectwise_baselines_v1/processed/`
#
# Note: "run_subjectwise_baselines_v1/processed/" is available at:
# https://drive.google.com/drive/folders/1K4J-jqfWtoG7njM21aOgc2Ct7QgYf4W9?usp=sharing
#
# ## Output philosophy
# Only a small set of result files is written so analysis can be done elsewhere if needed.
# %%
!pip install -q pandas numpy scipy scikit-learn matplotlib openpyxl torch
# %%
from google.colab import drive
drive.mount('/content/drive')
# %% [markdown]
# ## Imports and configuration
# %%
import os
import re
import json
import math
import random
import warnings
from pathlib import Path
from itertools import combinations
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy.signal import coherence
from sklearn.base import clone
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import GroupShuffleSplit
from sklearn.metrics import (
accuracy_score, balanced_accuracy_score, f1_score, matthews_corrcoef,
confusion_matrix, roc_auc_score, precision_score, recall_score
)
from sklearn.ensemble import ExtraTreesClassifier
from sklearn.neural_network import MLPClassifier
import torch
import torch.nn as nn
from torch.utils.data import TensorDataset, DataLoader
warnings.filterwarnings("ignore")
pd.set_option("display.max_columns", 200)
plt.rcParams.update({
"font.size": 11,
"axes.titlesize": 15,
"axes.labelsize": 13,
"xtick.labelsize": 11,
"ytick.labelsize": 11,
"legend.fontsize": 11,
"figure.dpi": 150,
"savefig.dpi": 300,
"axes.spines.top": False,
"axes.spines.right": False,
})
# %%
# =========================
# USER CONFIGURATION
# =========================
BASE_DIR = Path("/content/drive/MyDrive/...")
INPUT_DIR = BASE_DIR / "run_subjectwise_baselines_v1" / "processed"
RUN_DIR = BASE_DIR / "REVISION/run_final_public_notebook_v1"
RUN_DIR.mkdir(parents=True, exist_ok=True)
# Output files
FINAL_REPEAT_METRICS = RUN_DIR / "final_repeat_metrics.csv"
FINAL_SUMMARY = RUN_DIR / "final_summary.csv"
FINAL_SUBJECT_PREDICTIONS = RUN_DIR / "final_subject_predictions.csv"
FINAL_TOP_FEATURES = RUN_DIR / "final_top_features.csv"
FINAL_MANUSCRIPT_TABLE = RUN_DIR / "final_manuscript_table.csv"
FINAL_MANUSCRIPT_TABLE_XLSX = RUN_DIR / "final_manuscript_table.xlsx"
FINAL_FEATURE_SELECTION = RUN_DIR / "final_feature_selection.csv"
FINAL_FEATURE_SELECTION_SUMMARY = RUN_DIR / "final_feature_selection_summary.csv"
FIG_BASELINES = RUN_DIR / "fig_baselines_8ch.png"
FIG_FUSION = RUN_DIR / "fig_fusion_ablation.png"
FIG_IMPORTANCE = RUN_DIR / "fig_top_feature_importance.png"
FIG_FEATURE_SELECTION = RUN_DIR / "fig_feature_selection.png"
MANIFEST_PATH = RUN_DIR / "manifest.json"
RANDOM_SEED = 42
N_REPEATS = 10
TEST_SIZE_SUBJECT_FRACTION = 0.20
VAL_SIZE_FROM_TRAIN_SUBJECTS = 0.20
BATCH_SIZE = 64
CNN_EPOCHS = 25
CNN_LR = 1e-3
EARLY_STOPPING_PATIENCE = 5
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
FULL_CHANNELS = [
"Fp1", "Fp2", "F7", "F3", "Fz", "F4", "F8",
"T7", "C3", "Cz", "C4", "T8",
"P7", "P3", "Pz", "P4", "P8", "O1", "O2"
]
CHANNELS_8CH = ["Fp1", "Fp2", "F3", "F4", "C3", "C4", "P3", "P4"]
CONNECTIVITY_BANDS = {
"alpha": (8.0, 12.0),
"beta": (12.0, 30.0),
}
# Feature-selection ablation grid (top-K from Gini importance)
K_GRID = [5, 10, 15, 20, 30, 50, 70, 90]
# Bootstrap configuration for 95% CIs of the mean across repeats
BOOTSTRAP_N = 10000
BOOTSTRAP_CI = 95
BOOTSTRAP_SEED = 42
# %% [markdown]
# ## Helper functions
# %%
def set_all_seeds(seed: int):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
def specificity_from_cm(cm: np.ndarray) -> float:
if cm.shape != (2, 2):
return np.nan
tn, fp, fn, tp = cm.ravel()
return tn / (tn + fp) if (tn + fp) > 0 else np.nan
def safe_auc(y_true, y_prob) -> float:
try:
if len(np.unique(y_true)) < 2:
return np.nan
return roc_auc_score(y_true, y_prob)
except Exception:
return np.nan
def compute_metrics(y_true, y_pred, y_prob) -> dict:
cm = confusion_matrix(y_true, y_pred, labels=[0, 1])
sens = recall_score(y_true, y_pred, zero_division=0)
spec = specificity_from_cm(cm)
prec = precision_score(y_true, y_pred, zero_division=0)
return {
"accuracy": accuracy_score(y_true, y_pred),
"balanced_accuracy": balanced_accuracy_score(y_true, y_pred),
"precision": prec,
"recall_sensitivity": sens,
"specificity": spec,
"f1": f1_score(y_true, y_pred, zero_division=0),
"mcc": matthews_corrcoef(y_true, y_pred) if len(np.unique(y_true)) > 1 and len(np.unique(y_pred)) > 1 else np.nan,
"auroc": safe_auc(y_true, y_prob),
"tn": int(cm[0, 0]),
"fp": int(cm[0, 1]),
"fn": int(cm[1, 0]),
"tp": int(cm[1, 1]),
}
def aggregate_subject_predictions(df_pred: pd.DataFrame, subject_col: str = "subject_key") -> pd.DataFrame:
agg = df_pred.groupby(subject_col).agg(
y_true=("y_true", "first"),
prob=("prob", "mean"),
n_segments=("prob", "size"),
label=("label", "first"),
condition=("condition", lambda s: "|".join(sorted(pd.unique(s.astype(str))))),
subject_id=("subject_id", "first"),
).reset_index()
agg["y_pred"] = (agg["prob"] >= 0.5).astype(int)
return agg
def stratified_subject_holdout(subject_df: pd.DataFrame, test_fraction=0.2, seed=42):
rng = np.random.RandomState(seed)
train_subjects, test_subjects = [], []
for label_value in sorted(subject_df["y"].unique()):
sub = subject_df[subject_df["y"] == label_value]["subject_key"].tolist()
sub = list(sub)
rng.shuffle(sub)
n_test = max(1, int(round(len(sub) * test_fraction)))
test_subjects.extend(sub[:n_test])
train_subjects.extend(sub[n_test:])
return sorted(train_subjects), sorted(test_subjects)
def make_subject_label_table(df: pd.DataFrame) -> pd.DataFrame:
subj = df.groupby("subject_key").agg(
y=("label", lambda s: 1 if s.iloc[0] == "MDD" else 0),
label=("label", "first"),
subject_id=("subject_id", "first")
).reset_index()
return subj
def get_spectral_feature_columns(feature_table: pd.DataFrame, selected_channels: list) -> list:
meta_cols = {"file_name", "subject_id", "subject_key", "label", "condition"}
cols = [c for c in feature_table.columns if c not in meta_cols]
base_cols = []
asym_cols = []
for c in cols:
if "__ASYM_" in c:
asym_cols.append(c)
continue
parts = c.split("__")
if len(parts) != 2:
continue
prefix, ch = parts
if ch not in selected_channels:
continue
if prefix.endswith("_logabs") or prefix.endswith("_rel"):
base_cols.append(c)
keep_asym = []
for c in asym_cols:
m = re.search(r"__ASYM_([A-Za-z0-9]+)_([A-Za-z0-9]+)$", c)
if m and m.group(1) in selected_channels and m.group(2) in selected_channels:
keep_asym.append(c)
return sorted(base_cols + keep_asym)
def build_connectivity_features(X_raw_8ch, sfreq, channel_names, bands):
edge_pairs = list(combinations(range(len(channel_names)), 2))
edge_names = []
rows = []
for i, j in edge_pairs:
for band_name in bands.keys():
edge_names.append(f"coh_{band_name}__{channel_names[i]}__{channel_names[j]}")
for seg in X_raw_8ch:
feats = []
for i, j in edge_pairs:
x = seg[i]
y = seg[j]
f, cxy = coherence(x, y, fs=sfreq, nperseg=min(len(x), 512))
for band_name, (fmin, fmax) in bands.items():
mask = (f >= fmin) & (f < fmax)
val = float(np.mean(cxy[mask])) if np.any(mask) else 0.0
feats.append(val)
rows.append(feats)
return pd.DataFrame(rows, columns=edge_names)
def fit_predict_prob(model, X_train, y_train, X_test):
fitted = clone(model)
fitted.fit(X_train, y_train)
if hasattr(fitted, "predict_proba"):
prob = fitted.predict_proba(X_test)[:, 1]
else:
score = fitted.decision_function(X_test)
prob = 1 / (1 + np.exp(-score))
pred = (prob >= 0.5).astype(int)
return fitted, prob, pred
# ---------------------------------------------------------------------
# Deep-learning baselines: Compact 1D CNN, EEGNet, ShallowConvNet
# ---------------------------------------------------------------------
class CompactEEGCNN(nn.Module):
"""
Compact 1D CNN baseline operating directly on (B, C, T) raw segments.
"""
def __init__(self, n_channels: int, n_samples: int):
super().__init__()
self.net = nn.Sequential(
nn.Conv1d(n_channels, 32, kernel_size=7, padding=3),
nn.BatchNorm1d(32),
nn.ReLU(),
nn.MaxPool1d(2),
nn.Conv1d(32, 64, kernel_size=5, padding=2),
nn.BatchNorm1d(64),
nn.ReLU(),
nn.MaxPool1d(2),
nn.Conv1d(64, 128, kernel_size=5, padding=2),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.AdaptiveAvgPool1d(1)
)
self.head = nn.Sequential(
nn.Flatten(),
nn.Dropout(0.3),
nn.Linear(128, 64),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(64, 1)
)
def forward(self, x):
x = self.net(x)
x = self.head(x)
return x.squeeze(1)
class EEGNet(nn.Module):
"""
EEGNet (Lawhern et al., 2018). Compact 8-channel adaptation in PyTorch.
Inputs are shaped (B, C, T) and lifted to (B, 1, C, T) internally.
Notes
-----
F1 is the number of temporal filters in Block 1, D is the depth multiplier
for the depthwise spatial convolution, and F2 is the number of pointwise
filters in the separable convolution of Block 2. The temporal kernel
length defaults to half the sampling rate.
"""
def __init__(self, n_channels, n_samples, sfreq=256.0,
F1=8, D=2, F2=16, dropout=0.5, kern_length=None):
super().__init__()
if kern_length is None:
kern_length = max(8, int(sfreq // 2))
# Block 1: temporal convolution
self.block1 = nn.Sequential(
nn.Conv2d(1, F1, (1, kern_length),
padding=(0, kern_length // 2), bias=False),
nn.BatchNorm2d(F1),
)
# Block 1 cont: depthwise spatial convolution across channels
self.depthwise = nn.Sequential(
nn.Conv2d(F1, F1 * D, (n_channels, 1), groups=F1, bias=False),
nn.BatchNorm2d(F1 * D),
nn.ELU(),
nn.AvgPool2d((1, 4)),
nn.Dropout(dropout),
)
# Block 2: separable convolution (depthwise + pointwise)
self.separable = nn.Sequential(
nn.Conv2d(F1 * D, F1 * D, (1, 16),
padding=(0, 8), groups=F1 * D, bias=False),
nn.Conv2d(F1 * D, F2, (1, 1), bias=False),
nn.BatchNorm2d(F2),
nn.ELU(),
nn.AvgPool2d((1, 8)),
nn.Dropout(dropout),
)
with torch.no_grad():
x = torch.zeros(1, 1, n_channels, n_samples)
x = self.block1(x); x = self.depthwise(x); x = self.separable(x)
n_flat = x.numel()
self.head = nn.Linear(n_flat, 1)
def forward(self, x):
x = x.unsqueeze(1)
x = self.block1(x)
x = self.depthwise(x)
x = self.separable(x)
x = x.flatten(start_dim=1)
x = self.head(x)
return x.squeeze(1)
class ShallowConvNet(nn.Module):
"""
ShallowConvNet (Schirrmeister et al., 2017).
Temporal convolution, spatial convolution across channels, square
activation, mean pooling, log activation, dropout, and a linear head.
Inputs are shaped (B, C, T) and lifted to (B, 1, C, T) internally.
"""
def __init__(self, n_channels, n_samples,
n_filters_time=40, filter_time_length=25,
n_filters_spat=40, pool_time_length=75, pool_time_stride=15,
dropout=0.5):
super().__init__()
self.temporal = nn.Conv2d(1, n_filters_time, (1, filter_time_length), bias=False)
self.spatial = nn.Conv2d(n_filters_time, n_filters_spat, (n_channels, 1), bias=False)
self.bn = nn.BatchNorm2d(n_filters_spat)
self.pool = nn.AvgPool2d((1, pool_time_length), stride=(1, pool_time_stride))
self.dropout = nn.Dropout(dropout)
with torch.no_grad():
x = torch.zeros(1, 1, n_channels, n_samples)
x = self.temporal(x)
x = self.spatial(x)
x = self.bn(x)
x = x * x
x = self.pool(x)
x = torch.log(torch.clamp(x, min=1e-7, max=1e7))
n_flat = x.numel()
self.head = nn.Linear(n_flat, 1)
def forward(self, x):
x = x.unsqueeze(1)
x = self.temporal(x)
x = self.spatial(x)
x = self.bn(x)
x = x * x
x = self.pool(x)
x = torch.log(torch.clamp(x, min=1e-7, max=1e7))
x = self.dropout(x)
x = x.flatten(start_dim=1)
x = self.head(x)
return x.squeeze(1)
def fit_predict_torch_model(
model_factory, X_train, y_train, groups_train, X_test, seed,
val_fraction=0.2, batch_size=64, epochs=25, lr=1e-3, patience=5, device="cpu"
):
"""
Generic subject-wise trainer for any PyTorch (B, C, T) -> logit model.
model_factory(n_channels, n_samples) -> nn.Module returning a single
pre-sigmoid output per segment. Validation subjects are drawn from the
training partition with GroupShuffleSplit so that no subject appears in
both training and validation. Per-channel z-score normalisation
statistics are computed on the inner training split only.
"""
set_all_seeds(seed)
train_meta = pd.DataFrame({"subject_key": groups_train, "y": y_train}).drop_duplicates()
gss = GroupShuffleSplit(n_splits=1, test_size=val_fraction, random_state=seed)
tr_sub_idx, va_sub_idx = next(gss.split(train_meta[["y"]], train_meta["y"], groups=train_meta["subject_key"]))
tr_subjects = set(train_meta.iloc[tr_sub_idx]["subject_key"])
va_subjects = set(train_meta.iloc[va_sub_idx]["subject_key"])
tr_mask = np.array([g in tr_subjects for g in groups_train])
va_mask = np.array([g in va_subjects for g in groups_train])
X_tr = X_train[tr_mask]
y_tr = y_train[tr_mask]
X_va = X_train[va_mask]
y_va = y_train[va_mask]
mean = X_tr.mean(axis=(0, 2), keepdims=True)
std = X_tr.std(axis=(0, 2), keepdims=True) + 1e-6
X_tr = (X_tr - mean) / std
X_va = (X_va - mean) / std
X_te = (X_test - mean) / std
tr_ds = TensorDataset(torch.tensor(X_tr, dtype=torch.float32), torch.tensor(y_tr, dtype=torch.float32))
va_ds = TensorDataset(torch.tensor(X_va, dtype=torch.float32), torch.tensor(y_va, dtype=torch.float32))
te_x = torch.tensor(X_te, dtype=torch.float32)
tr_loader = DataLoader(tr_ds, batch_size=batch_size, shuffle=True)
va_loader = DataLoader(va_ds, batch_size=batch_size, shuffle=False)
n_channels = X_train.shape[1]
n_samples = X_train.shape[2]
model = model_factory(n_channels, n_samples).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
criterion = nn.BCEWithLogitsLoss()
best_state = None
best_val = np.inf
bad_epochs = 0
for _ in range(epochs):
model.train()
for xb, yb in tr_loader:
xb = xb.to(device); yb = yb.to(device)
optimizer.zero_grad()
logits = model(xb)
loss = criterion(logits, yb)
loss.backward()
optimizer.step()
model.eval()
val_losses = []
with torch.no_grad():
for xb, yb in va_loader:
xb = xb.to(device); yb = yb.to(device)
logits = model(xb)
val_losses.append(criterion(logits, yb).item())
val_loss = float(np.mean(val_losses)) if val_losses else np.inf
if val_loss < best_val:
best_val = val_loss
best_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}
bad_epochs = 0
else:
bad_epochs += 1
if bad_epochs >= patience:
break
if best_state is not None:
model.load_state_dict(best_state)
model.eval()
with torch.no_grad():
logits = model(te_x.to(device)).cpu().numpy()
prob = 1 / (1 + np.exp(-logits))
pred = (prob >= 0.5).astype(int)
return prob, pred
# ---------------------------------------------------------------------
# Bootstrap CI helper
# ---------------------------------------------------------------------
def bootstrap_ci_mean(values, n_bootstrap=10000, ci=95, seed=42):
"""
Percentile bootstrap confidence interval of the mean.
Returns (mean, lower, upper). NaNs are dropped before resampling. If
fewer than two finite values remain, the bounds are returned as NaN.
"""
values = np.asarray(values, dtype=float)
values = values[~np.isnan(values)]
if len(values) < 2:
m = float(np.mean(values)) if len(values) else float("nan")
return m, float("nan"), float("nan")
rng = np.random.RandomState(seed)
n = len(values)
boots = rng.choice(values, size=(n_bootstrap, n), replace=True).mean(axis=1)
alpha = (100 - ci) / 2.0
return float(np.mean(values)), float(np.percentile(boots, alpha)), float(np.percentile(boots, 100 - alpha))
def make_public_label(model_name: str) -> str:
mapping = {
"extratrees_8ch_features": "Extra Trees (features)",
"mlp_8ch_features": "MLP (features)",
"cnn1d_8ch_raw": "1D CNN (raw 8ch)",
"eegnet_8ch_raw": "EEGNet (raw 8ch)",
"shallowconvnet_8ch_raw": "ShallowConvNet (raw 8ch)",
"extratrees_spectral": "ET (spectral)",
"extratrees_connectivity": "ET (connectivity)",
"extratrees_fusion": "ET (fusion)",
"mlp_fusion": "MLP (fusion)",
}
return mapping.get(model_name, model_name)
# %% [markdown]
# ## Load processed data and create the corrected subject key
# %%
feature_table = pd.read_csv(INPUT_DIR / "feature_table.csv")
X_segments = np.load(INPUT_DIR / "X_segments.npy")
y_segments = np.load(INPUT_DIR / "y.npy")
groups_segments = np.load(INPUT_DIR / "groups.npy", allow_pickle=True)
conditions_segments = np.load(INPUT_DIR / "conditions.npy", allow_pickle=True)
feature_table["subject_id"] = feature_table["subject_id"].astype(str)
feature_table["label"] = feature_table["label"].astype(str)
feature_table["condition"] = feature_table["condition"].astype(str)
feature_table["subject_key"] = feature_table["label"] + "_" + feature_table["subject_id"]
group_subject_ids = pd.Series(groups_segments.astype(str))
group_labels = pd.Series(np.where(y_segments == 1, "MDD", "H"))
group_subject_keys = (group_labels + "_" + group_subject_ids).values
subject_df = make_subject_label_table(feature_table)
print("feature_table:", feature_table.shape)
print("X_segments:", X_segments.shape)
print("Unique subject_key count:", len(subject_df))
print("Subjects by class:")
print(subject_df["label"].value_counts())
# %% [markdown]
# ## Prepare feature matrices and raw 8-channel tensor
# %%
channel_idx = [FULL_CHANNELS.index(ch) for ch in CHANNELS_8CH]
X_raw_8ch = X_segments[:, channel_idx, :]
sfreq = 256.0
spectral_cols = get_spectral_feature_columns(feature_table, CHANNELS_8CH)
X_spectral = feature_table[spectral_cols].copy()
X_connectivity = build_connectivity_features(
X_raw_8ch=X_raw_8ch,
sfreq=sfreq,
channel_names=CHANNELS_8CH,
bands=CONNECTIVITY_BANDS
)
X_fusion = pd.concat([X_spectral.reset_index(drop=True), X_connectivity.reset_index(drop=True)], axis=1)
y_all = feature_table["label"].map({"H": 0, "MDD": 1}).astype(int).values
subject_keys_all = feature_table["subject_key"].astype(str).values
subject_ids_all = feature_table["subject_id"].astype(str).values
conditions_all = feature_table["condition"].astype(str).values
print("Spectral features:", X_spectral.shape)
print("Connectivity features:", X_connectivity.shape)
print("Fusion features:", X_fusion.shape)
print("Raw 8ch tensor:", X_raw_8ch.shape)
# %% [markdown]
# ## Define the experimental models
# %%
compact_models = {
"extratrees_8ch_features": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300,
class_weight="balanced",
random_state=RANDOM_SEED,
n_jobs=-1
))
]),
"mlp_8ch_features": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(128, 64),
activation="relu",
alpha=1e-4,
batch_size=64,
learning_rate_init=1e-3,
max_iter=200,
early_stopping=True,
validation_fraction=0.15,
random_state=RANDOM_SEED
))
]),
}
# Deep raw-signal baselines: each entry maps a model id to a factory taking
# (n_channels, n_samples) and returning an nn.Module. All are trained with
# fit_predict_torch_model under the same subject-wise inner validation split.
deep_models = {
"cnn1d_8ch_raw": (
lambda n_channels, n_samples: CompactEEGCNN(n_channels, n_samples)
),
"eegnet_8ch_raw": (
lambda n_channels, n_samples: EEGNet(
n_channels=n_channels, n_samples=n_samples, sfreq=sfreq,
F1=8, D=2, F2=16, dropout=0.5
)
),
"shallowconvnet_8ch_raw": (
lambda n_channels, n_samples: ShallowConvNet(
n_channels=n_channels, n_samples=n_samples,
n_filters_time=40, filter_time_length=25,
n_filters_spat=40, pool_time_length=75, pool_time_stride=15,
dropout=0.5
)
),
}
fusion_models = {
"extratrees_spectral": {
"X": X_spectral,
"model": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300,
class_weight="balanced",
random_state=RANDOM_SEED,
n_jobs=-1
))
])
},
"extratrees_connectivity": {
"X": X_connectivity,
"model": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300,
class_weight="balanced",
random_state=RANDOM_SEED,
n_jobs=-1
))
])
},
"extratrees_fusion": {
"X": X_fusion,
"model": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300,
class_weight="balanced",
random_state=RANDOM_SEED,
n_jobs=-1
))
])
},
"mlp_fusion": {
"X": X_fusion,
"model": Pipeline([
("imputer", SimpleImputer(strategy="median")),
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(128, 64),
activation="relu",
alpha=1e-4,
batch_size=64,
learning_rate_init=1e-3,
max_iter=200,
early_stopping=True,
validation_fraction=0.15,
random_state=RANDOM_SEED
))
])
},
}
# %% [markdown]
# ## Run all experiments under repeated subject-wise holdout
# %%
all_metric_rows = []
all_subject_pred_rows = []
feature_importance_rows = []
for repeat_idx in range(N_REPEATS):
seed = RANDOM_SEED + repeat_idx
set_all_seeds(seed)
train_subjects, test_subjects = stratified_subject_holdout(
subject_df,
test_fraction=TEST_SIZE_SUBJECT_FRACTION,
seed=seed
)
train_mask = np.isin(subject_keys_all, train_subjects)
test_mask = np.isin(subject_keys_all, test_subjects)
y_train = y_all[train_mask]
y_test = y_all[test_mask]
g_train = subject_keys_all[train_mask]
g_test = subject_keys_all[test_mask]
sid_test = subject_ids_all[test_mask]
cond_test = conditions_all[test_mask]
label_test = np.where(y_test == 1, "MDD", "H")
print(f"===== Repeat {repeat_idx+1}/{N_REPEATS} | seed={seed} =====")
tmp = pd.DataFrame({"subject_key": train_subjects + test_subjects, "split": ["train"]*len(train_subjects)+["test"]*len(test_subjects)})
tmp = tmp.merge(subject_df[["subject_key", "label"]], on="subject_key", how="left")
print(tmp.groupby(["split", "label"]).size())
# -----------------------------------------------------------------
# Experiment 1a: tabular feature-based compact models
# -----------------------------------------------------------------
Xf_train = X_spectral.loc[train_mask].reset_index(drop=True)
Xf_test = X_spectral.loc[test_mask].reset_index(drop=True)
for model_name, model in compact_models.items():
fitted, prob, pred = fit_predict_prob(model, Xf_train, y_train, Xf_test)
seg_df = pd.DataFrame({
"repeat": repeat_idx, "seed": seed, "model": model_name,
"subject_key": g_test, "subject_id": sid_test, "label": label_test,
"condition": cond_test, "y_true": y_test, "prob": prob, "y_pred": pred,
})
subj_df = aggregate_subject_predictions(seg_df, subject_col="subject_key")
subj_df["repeat"] = repeat_idx
subj_df["seed"] = seed
subj_df["model"] = model_name
all_subject_pred_rows.append(subj_df)
metrics = compute_metrics(subj_df["y_true"].values, subj_df["y_pred"].values, subj_df["prob"].values)
metrics.update({"repeat": repeat_idx, "seed": seed, "model": model_name, "experiment": "compact_baselines", "n_subjects": len(subj_df), "n_segments": len(seg_df)})
all_metric_rows.append(metrics)
if model_name == "extratrees_8ch_features":
clf = fitted.named_steps["clf"]
feat_names = list(Xf_train.columns)
importances = clf.feature_importances_
top_idx = np.argsort(importances)[::-1][:15]
for rank, idx in enumerate(top_idx, start=1):
feature_importance_rows.append({
"repeat": repeat_idx, "model": model_name, "rank": rank,
"feature": feat_names[idx], "importance": float(importances[idx]),
"family": "spectral",
})
# -----------------------------------------------------------------
# Experiment 1b: deep raw-signal baselines (1D CNN, EEGNet, ShallowConvNet)
# -----------------------------------------------------------------
Xr_train = X_raw_8ch[train_mask]
Xr_test = X_raw_8ch[test_mask]
for model_name, factory in deep_models.items():
prob_dl, pred_dl = fit_predict_torch_model(
model_factory=factory,
X_train=Xr_train, y_train=y_train, groups_train=g_train, X_test=Xr_test,
seed=seed, val_fraction=VAL_SIZE_FROM_TRAIN_SUBJECTS, batch_size=BATCH_SIZE,
epochs=CNN_EPOCHS, lr=CNN_LR, patience=EARLY_STOPPING_PATIENCE, device=DEVICE,
)
seg_df = pd.DataFrame({
"repeat": repeat_idx, "seed": seed, "model": model_name,
"subject_key": g_test, "subject_id": sid_test, "label": label_test,
"condition": cond_test, "y_true": y_test, "prob": prob_dl, "y_pred": pred_dl,
})
subj_df = aggregate_subject_predictions(seg_df, subject_col="subject_key")
subj_df["repeat"] = repeat_idx
subj_df["seed"] = seed
subj_df["model"] = model_name
all_subject_pred_rows.append(subj_df)
metrics = compute_metrics(subj_df["y_true"].values, subj_df["y_pred"].values, subj_df["prob"].values)
metrics.update({"repeat": repeat_idx, "seed": seed, "model": model_name, "experiment": "compact_baselines", "n_subjects": len(subj_df), "n_segments": len(seg_df)})
all_metric_rows.append(metrics)
# -----------------------------------------------------------------
# Experiment 2: connectivity ablation (unchanged)
# -----------------------------------------------------------------
for model_name, spec in fusion_models.items():
X_df = spec["X"]
model = spec["model"]
X_train_df = X_df.loc[train_mask].reset_index(drop=True)
X_test_df = X_df.loc[test_mask].reset_index(drop=True)
fitted, prob, pred = fit_predict_prob(model, X_train_df, y_train, X_test_df)
seg_df = pd.DataFrame({
"repeat": repeat_idx, "seed": seed, "model": model_name,
"subject_key": g_test, "subject_id": sid_test, "label": label_test,
"condition": cond_test, "y_true": y_test, "prob": prob, "y_pred": pred,
})
subj_df = aggregate_subject_predictions(seg_df, subject_col="subject_key")
subj_df["repeat"] = repeat_idx
subj_df["seed"] = seed
subj_df["model"] = model_name
all_subject_pred_rows.append(subj_df)
metrics = compute_metrics(subj_df["y_true"].values, subj_df["y_pred"].values, subj_df["prob"].values)
metrics.update({"repeat": repeat_idx, "seed": seed, "model": model_name, "experiment": "fusion_ablation", "n_subjects": len(subj_df), "n_segments": len(seg_df)})
all_metric_rows.append(metrics)
if model_name.startswith("extratrees"):
clf = fitted.named_steps["clf"]
feat_names = list(X_df.columns)
importances = clf.feature_importances_
top_idx = np.argsort(importances)[::-1][:15]
family = "spectral" if model_name == "extratrees_spectral" else ("connectivity" if model_name == "extratrees_connectivity" else "fusion")
for rank, idx in enumerate(top_idx, start=1):
feature_importance_rows.append({
"repeat": repeat_idx, "model": model_name, "rank": rank,
"feature": feat_names[idx], "importance": float(importances[idx]),
"family": family,
})
# %% [markdown]
# ## Experiment 3 — Spectral feature-selection ablation
#
# To test for potential overfitting from a 90-dimensional spectral feature set, we run a top-$K$ Gini-importance ablation. For each repeat we fit Extra Trees on the full 90-feature set, rank the columns by importance, then refit and re-evaluate Extra Trees on the top $K \in \{5, 10, 15, 20, 30, 50, 70, 90\}$ features under the same subject-wise test split. Subject-level metrics are reported as mean and 95% bootstrap CI across the 10 repeats.
# %%
fs_metric_rows = []
for repeat_idx in range(N_REPEATS):
seed = RANDOM_SEED + repeat_idx
set_all_seeds(seed)
train_subjects, test_subjects = stratified_subject_holdout(
subject_df, test_fraction=TEST_SIZE_SUBJECT_FRACTION, seed=seed
)
train_mask = np.isin(subject_keys_all, train_subjects)
test_mask = np.isin(subject_keys_all, test_subjects)
y_train = y_all[train_mask]
y_test = y_all[test_mask]
g_test = subject_keys_all[test_mask]
sid_test = subject_ids_all[test_mask]
cond_test = conditions_all[test_mask]
label_test = np.where(y_test == 1, "MDD", "H")
Xf_train = X_spectral.loc[train_mask].reset_index(drop=True)
Xf_test = X_spectral.loc[test_mask].reset_index(drop=True)
# Fit on the full feature set to obtain Gini-importance ranking
full_pipe = Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300, class_weight="balanced",
random_state=seed, n_jobs=-1
))
])
full_pipe.fit(Xf_train, y_train)
importances = full_pipe.named_steps["clf"].feature_importances_
feature_names = list(Xf_train.columns)
n_total = len(feature_names)
sorted_idx = np.argsort(importances)[::-1]
for K in K_GRID:
k_use = min(K, n_total)
keep_idx = sorted_idx[:k_use]
keep_cols = [feature_names[i] for i in keep_idx]
Xs_train = Xf_train[keep_cols]
Xs_test = Xf_test[keep_cols]
sub_pipe = Pipeline([
("imputer", SimpleImputer(strategy="median")),
("clf", ExtraTreesClassifier(
n_estimators=300, class_weight="balanced",
random_state=seed, n_jobs=-1
))
])
_, prob, pred = fit_predict_prob(sub_pipe, Xs_train, y_train, Xs_test)
seg_df = pd.DataFrame({
"repeat": repeat_idx, "seed": seed,
"model": f"extratrees_topk_{K}",
"subject_key": g_test, "subject_id": sid_test,
"label": label_test, "condition": cond_test,
"y_true": y_test, "prob": prob, "y_pred": pred,
})
subj_df = aggregate_subject_predictions(seg_df, subject_col="subject_key")
m = compute_metrics(subj_df["y_true"].values, subj_df["y_pred"].values, subj_df["prob"].values)
m.update({
"repeat": repeat_idx, "seed": seed,
"K": K, "n_features_used": k_use,
"experiment": "feature_selection",
"model": f"extratrees_topk_{K}",
"n_subjects": len(subj_df), "n_segments": len(seg_df),
})
fs_metric_rows.append(m)
print(f"FS ablation repeat {repeat_idx+1}/{N_REPEATS} done")
fs_metrics_df = pd.DataFrame(fs_metric_rows)
fs_metrics_df.to_csv(FINAL_FEATURE_SELECTION, index=False)
# Per-K summary with bootstrap 95% CI of the mean
fs_summary_rows = []
for K, grp in fs_metrics_df.groupby("K"):
rec = {"K": int(K), "n_features_used": int(grp["n_features_used"].iloc[0])}
for raw_name, out_name in [
("balanced_accuracy", "balanced_accuracy"),
("auroc", "auroc"),
("recall_sensitivity", "sensitivity"),
("specificity", "specificity"),
("precision", "precision"),
("f1", "f1"),
("mcc", "mcc"),
]:
vals = grp[raw_name].values
m, lo, hi = bootstrap_ci_mean(vals, n_bootstrap=BOOTSTRAP_N, ci=BOOTSTRAP_CI, seed=BOOTSTRAP_SEED)
rec[f"{out_name}_mean"] = m
rec[f"{out_name}_std"] = float(np.nanstd(vals, ddof=1)) if np.sum(~np.isnan(vals)) > 1 else float("nan")
rec[f"{out_name}_ci_lower"] = lo
rec[f"{out_name}_ci_upper"] = hi
fs_summary_rows.append(rec)
fs_summary_df = pd.DataFrame(fs_summary_rows).sort_values("K").reset_index(drop=True)
fs_summary_df.to_csv(FINAL_FEATURE_SELECTION_SUMMARY, index=False)
print("Saved:")
print(FINAL_FEATURE_SELECTION)
print(FINAL_FEATURE_SELECTION_SUMMARY)
display(fs_summary_df[[
"K", "balanced_accuracy_mean", "balanced_accuracy_ci_lower", "balanced_accuracy_ci_upper",
"auroc_mean", "auroc_ci_lower", "auroc_ci_upper",
]])
# %% [markdown]
# ## Save tables
# %%
repeat_metrics_df = pd.DataFrame(all_metric_rows)
subject_predictions_df = pd.concat(all_subject_pred_rows, axis=0, ignore_index=True) if all_subject_pred_rows else pd.DataFrame()
top_features_df = pd.DataFrame(feature_importance_rows)
# Mean and SD across the 10 repeats per (experiment, model)
summary_df = (
repeat_metrics_df
.groupby(["experiment", "model"], as_index=False)
.agg(
balanced_accuracy_mean=("balanced_accuracy", "mean"),
balanced_accuracy_std=("balanced_accuracy", "std"),
auroc_mean=("auroc", "mean"),
auroc_std=("auroc", "std"),
sensitivity_mean=("recall_sensitivity", "mean"),
sensitivity_std=("recall_sensitivity", "std"),
specificity_mean=("specificity", "mean"),
specificity_std=("specificity", "std"),
precision_mean=("precision", "mean"),
precision_std=("precision", "std"),
f1_mean=("f1", "mean"),
f1_std=("f1", "std"),
mcc_mean=("mcc", "mean"),
mcc_std=("mcc", "std"),
n_subjects_mean=("n_subjects", "mean"),
)
.sort_values(["experiment", "balanced_accuracy_mean", "auroc_mean"], ascending=[True, False, False])
.reset_index(drop=True)
)
summary_df["public_label"] = summary_df["model"].map(make_public_label)
# Bootstrap 95% CI of the mean across the 10 repeats, per (experiment, model)
metric_columns = {
"balanced_accuracy": "balanced_accuracy",
"auroc": "auroc",
"recall_sensitivity": "sensitivity",
"specificity": "specificity",
"precision": "precision",
"f1": "f1",
"mcc": "mcc",
}
ci_records = []
for (exp_name, model_name), grp in repeat_metrics_df.groupby(["experiment", "model"]):
rec = {"experiment": exp_name, "model": model_name}
for raw_name, out_name in metric_columns.items():
vals = grp[raw_name].values
m, lo, hi = bootstrap_ci_mean(vals, n_bootstrap=BOOTSTRAP_N, ci=BOOTSTRAP_CI, seed=BOOTSTRAP_SEED)
rec[f"{out_name}_ci_lower"] = lo
rec[f"{out_name}_ci_upper"] = hi
ci_records.append(rec)
ci_df = pd.DataFrame(ci_records)
summary_df = summary_df.merge(ci_df, on=["experiment", "model"], how="left")
# Top features summary (unchanged)
if len(top_features_df):
top_features_summary = (
top_features_df
.groupby(["model", "family", "feature"], as_index=False)
.agg(
importance_mean=("importance", "mean"),
importance_count=("importance", "size"),
)
.sort_values(["model", "importance_mean"], ascending=[True, False])
.groupby("model", as_index=False)
.head(12)
.reset_index(drop=True)
)
else:
top_features_summary = pd.DataFrame()
# Manuscript table now includes both SD and 95% CI columns for the headline
# metrics. Editors and reviewers can pick whichever uncertainty representation
# they prefer when reading Tables 1 and 2.
manuscript_table = summary_df[[
"experiment", "public_label",
"balanced_accuracy_mean", "balanced_accuracy_std",
"balanced_accuracy_ci_lower", "balanced_accuracy_ci_upper",
"auroc_mean", "auroc_std",
"auroc_ci_lower", "auroc_ci_upper",
"sensitivity_mean", "sensitivity_std",
"specificity_mean", "specificity_std",
"precision_mean", "precision_std",
"f1_mean", "f1_std",
"mcc_mean", "mcc_std",
]].copy()
repeat_metrics_df.to_csv(FINAL_REPEAT_METRICS, index=False)
summary_df.to_csv(FINAL_SUMMARY, index=False)
subject_predictions_df.to_csv(FINAL_SUBJECT_PREDICTIONS, index=False)
top_features_summary.to_csv(FINAL_TOP_FEATURES, index=False)
manuscript_table.to_csv(FINAL_MANUSCRIPT_TABLE, index=False)
manuscript_table.to_excel(FINAL_MANUSCRIPT_TABLE_XLSX, index=False)
print("Saved:")
print(FINAL_REPEAT_METRICS)
print(FINAL_SUMMARY)
print(FINAL_SUBJECT_PREDICTIONS)
print(FINAL_TOP_FEATURES)
print(FINAL_MANUSCRIPT_TABLE)
print(FINAL_MANUSCRIPT_TABLE_XLSX)
display(manuscript_table)
# %% [markdown]
# ## Figure 1 — Compact 8-channel baselines
# %%
fig, ax = plt.subplots(figsize=(10.0, 4.8), constrained_layout=True)
df_plot = summary_df[summary_df["experiment"] == "compact_baselines"].copy()
order = [
"extratrees_8ch_features",
"mlp_8ch_features",
"cnn1d_8ch_raw",
"eegnet_8ch_raw",
"shallowconvnet_8ch_raw",
]
df_plot["order"] = df_plot["model"].map({m: i for i, m in enumerate(order)})
df_plot = df_plot.sort_values("order")
labels = df_plot["model"].map(make_public_label).tolist()
means = df_plot["balanced_accuracy_mean"].values
lower = df_plot["balanced_accuracy_ci_lower"].values
upper = df_plot["balanced_accuracy_ci_upper"].values
err_low = means - lower
err_high = upper - means
# Wong colour-blind-safe palette. Feature-based models use cool tones,
# raw-signal deep models use warm tones, and the strongest model (Extra
# Trees on spectral features) anchors the palette in deep bluish-green.
bar_colors = [
"#009E73", # Extra Trees (spectral) - winner, deep bluish-green
"#56B4E9", # MLP (spectral) - feature-based, sky blue
"#E69F00", # 1D CNN (raw) - raw-signal, orange
"#D55E00", # EEGNet (raw) - raw-signal, vermilion
"#CC79A7", # ShallowConvNet (raw) - raw-signal, reddish purple
]
edge_colors = ["#006B4F", "#2A7BB0", "#A06D00", "#933F00", "#8C4F75"]
ax.set_axisbelow(True)
bars = ax.bar(
range(len(labels)), means,
yerr=[err_low, err_high],
color=bar_colors, edgecolor=edge_colors, linewidth=0.9,
capsize=4, error_kw={"elinewidth": 1.2, "ecolor": "#333333"},
alpha=0.95,
)
for i, v in enumerate(means):
ax.text(i, upper[i] + 0.015, f"{v:.3f}", ha="center", va="bottom", fontsize=11)
ax.set_xticks(range(len(labels)))
ax.set_xticklabels(labels, rotation=12, ha="right")
ax.set_ylabel("Subject-level balanced accuracy")
ax.set_title("Compact 8-channel baselines (mean and 95% bootstrap CI)")
ax.set_ylim(0.0, 1.05)
ax.set_yticks(np.arange(0.0, 1.01, 0.2))
ax.grid(axis="y", linewidth=0.4, alpha=0.3)
fig.savefig(FIG_BASELINES, bbox_inches="tight")
plt.show()
print(FIG_BASELINES)
# %% [markdown]
# ## Figure 2 — Spectral vs connectivity vs fusion
# %%
fig, ax = plt.subplots(figsize=(8.8, 4.8), constrained_layout=True)
df_plot = summary_df[summary_df["experiment"] == "fusion_ablation"].copy()
order = ["extratrees_spectral", "extratrees_connectivity", "extratrees_fusion", "mlp_fusion"]
df_plot["order"] = df_plot["model"].map({m: i for i, m in enumerate(order)})
df_plot = df_plot.sort_values("order")
labels = df_plot["model"].map(make_public_label).tolist()
means = df_plot["balanced_accuracy_mean"].values
lower = df_plot["balanced_accuracy_ci_lower"].values
upper = df_plot["balanced_accuracy_ci_upper"].values
err_low = means - lower
err_high = upper - means
# Distinct colour per feature condition. The spectral bar uses the same
# green as the winning model in Figure 1, providing visual continuity
# across the two experiments. Connectivity gets a warm orange to mark its
# different feature family. Fusion gets a deep unifying blue. MLP fusion
# gets a complementary reddish purple.
bar_colors = [
"#009E73", # ET (spectral) - matches Fig 1 winner
"#E69F00", # ET (connectivity) - warm orange for connectivity family
"#0072B2", # ET (fusion) - deep blue for combined representation
"#CC79A7", # MLP (fusion) - reddish purple
]
edge_colors = ["#006B4F", "#A06D00", "#004F7C", "#8C4F75"]
ax.set_axisbelow(True)
bars = ax.bar(
range(len(labels)), means,
yerr=[err_low, err_high],
color=bar_colors, edgecolor=edge_colors, linewidth=0.9,
capsize=4, error_kw={"elinewidth": 1.2, "ecolor": "#333333"},
alpha=0.95,
)
for i, v in enumerate(means):
ax.text(i, upper[i] + 0.015, f"{v:.3f}", ha="center", va="bottom", fontsize=11)
ax.set_xticks(range(len(labels)))
ax.set_xticklabels(labels)
ax.set_ylabel("Subject-level balanced accuracy")
ax.set_title("Spectral vs connectivity vs fusion (mean and 95% bootstrap CI)")
ax.set_ylim(0.0, 1.05)
ax.set_yticks(np.arange(0.0, 1.01, 0.2))
ax.grid(axis="y", linewidth=0.4, alpha=0.3)
fig.savefig(FIG_FUSION, bbox_inches="tight")
plt.show()
print(FIG_FUSION)
# %% [markdown]
# ## Figure 3 — Top feature importance preview
# %%
fig, axs = plt.subplots(1, 2, figsize=(12.5, 4.8), constrained_layout=True)
# (a) Spectral winner. Sequential green ramp keeps the spectral colour
# family from Figures 1 and 2, with deeper colour for higher importance.
spec_df = top_features_summary[top_features_summary["model"] == "extratrees_8ch_features"].head(10).copy()
spec_df = spec_df.sort_values("importance_mean", ascending=True)
spec_colors = plt.cm.YlGn(np.linspace(0.45, 0.85, len(spec_df)))
axs[0].set_axisbelow(True)
axs[0].barh(spec_df["feature"], spec_df["importance_mean"],
color=spec_colors, edgecolor="#2D5F3E", linewidth=0.6)
axs[0].set_title("(a) Top spectral/asymmetry features")
axs[0].set_xlabel("Mean importance")
axs[0].grid(axis="x", linewidth=0.4, alpha=0.3)
# (b) Connectivity winner. Sequential orange ramp matches the
# connectivity colour family used for the connectivity bar in Figure 2.
conn_df = top_features_summary[top_features_summary["model"] == "extratrees_connectivity"].head(10).copy()
conn_df = conn_df.sort_values("importance_mean", ascending=True)
conn_colors = plt.cm.YlOrBr(np.linspace(0.40, 0.80, len(conn_df)))
axs[1].set_axisbelow(True)
axs[1].barh(conn_df["feature"], conn_df["importance_mean"],
color=conn_colors, edgecolor="#7A4F1F", linewidth=0.6)
axs[1].set_title("(b) Top connectivity features")
axs[1].set_xlabel("Mean importance")
axs[1].grid(axis="x", linewidth=0.4, alpha=0.3)
fig.savefig(FIG_IMPORTANCE, bbox_inches="tight")
plt.show()
print(FIG_IMPORTANCE)
# %% [markdown]
# ## Figure 4 — Spectral feature-selection ablation curve
#
# Subject-level balanced accuracy and 95% bootstrap CI as a function of the number of selected features. The full 90-feature baseline is shown as a horizontal reference. If accuracy plateaus well below $K=90$, the spectral representation is not over-parameterised; if it improves at smaller $K$, top-$K$ pruning offers a parsimonious alternative.
# %%
fig, ax = plt.subplots(figsize=(8.6, 4.8), constrained_layout=True)
xs = fs_summary_df["K"].values
means = fs_summary_df["balanced_accuracy_mean"].values
los = fs_summary_df["balanced_accuracy_ci_lower"].values
his = fs_summary_df["balanced_accuracy_ci_upper"].values
# Reuse the spectral-winner colour from Figure 1 since this figure
# describes the same Extra Trees on spectral features at varying K.
PRIMARY = "#009E73"
PRIMARY_DARK = "#006B4F"
ax.set_axisbelow(True)
ax.fill_between(xs, los, his, color=PRIMARY, alpha=0.18, label="95% bootstrap CI")
ax.plot(xs, means, "o-", color=PRIMARY, linewidth=2.0, markersize=7,
markerfacecolor="white", markeredgecolor=PRIMARY_DARK, markeredgewidth=1.8,
label="Top-$K$ Extra Trees mean")
# Reference line for the full 90-feature baseline
full_row = fs_summary_df[fs_summary_df["K"] == 90]
if len(full_row):
ref = float(full_row["balanced_accuracy_mean"].iloc[0])
ax.axhline(ref, linestyle="--", color="#555555", linewidth=1.0, alpha=0.7,
label=f"Full 90-feature baseline ({ref:.3f})")
for x, m, lo, hi in zip(xs, means, los, his):
ax.text(x, hi + 0.012, f"{m:.3f}", ha="center", va="bottom", fontsize=10)
ax.set_xlabel("Number of selected spectral and asymmetry features ($K$)")
ax.set_ylabel("Subject-level balanced accuracy")
ax.set_title("Spectral feature-selection ablation (Extra Trees, mean and 95% CI)")
ax.set_xticks(xs)
ymin = float(np.nanmin(los)) - 0.04
ymax = max(1.02, float(np.nanmax(his)) + 0.05)
ax.set_ylim(max(0.0, ymin), ymax)
ax.grid(axis="y", linewidth=0.4, alpha=0.3)
ax.legend(loc="lower right", frameon=False)
fig.savefig(FIG_FEATURE_SELECTION, bbox_inches="tight")
plt.show()
print(FIG_FEATURE_SELECTION)

EEG_Depression_Analysis_Public_Notebook.ipynb at commit 0a6888f, no license · at the source

Overview

Authors: Hassan Ugail¹, Newton Howard², Ali Ahmed Elmahmudi¹, Zied Mnasri¹

ORCID iDs: Hassan Ugail, Newton Howard, Ali Ahmed Elmahmudi, Zied Mnasri

Institutions: University of Bradford (United Kingdom); Howard Brain Sciences Foundation (United States)

Journal: Sensors (Basel, Switzerland), volume 26, issue 10, article 3065

Dates: received 9 April 2026; accepted 11 May 2026; published online 12 May 2026

Type: Research article · Language: English

License: CC BY

Identifiers: DOI 10.3390/s26103065 · PMID 42197874 · PMCID PMC13210619 · OpenAlex W7160913777

Open access: gold, a free copy (OpenAlex)

Status: code verified

Categories: EEG (modality), other (modality), human (organism), depression (population), cellular / molecular (subfield)

Methods: Spectral & time-frequency, Connectivity, Preprocessing, Statistics, Machine learning

Keywords: electroencephalography, major depressive disorder, subject-wise evaluation, data leakage, spectral features, frontal alpha asymmetry, beta-band power, inter-channel coherence, Extra Trees, wearable EEG

MeSH: Depression*, Electroencephalography*, Major Depressive Disorder*, Adult, Algorithms, Convolutional Neural Networks, Female, Humans, Male, Rest, Signal Processing, Computer-Assisted (* major topic)

Topic: EEG and Brain-Computer Interfaces (Cognitive Neuroscience, Neuroscience), according to OpenAlex

Funding: Howard Brain Sciences Foundation

Citations: not cited yet (Europe PMC); 34 references in the paper

Abstract

Major depressive disorder remains difficult to diagnose objectively, as routine assessment is still largely dependent on clinical interview and rating scales. Resting-state electroencephalography (EEG) is an attractive complementary modality because it is non-invasive, low-cost, and compatible with wearable sensing, but many reported EEG classification results are weakened by segment-level leakage and unclear subject identity handling. This study evaluates whether depression can be distinguished from healthy controls using a compact eight-channel resting-state EEG configuration under a strictly leakage-free subject-wise protocol. Using a widely used public EEG dataset, we first corrected a previously overlooked subject-identity ambiguity by constructing a class-aware composite key, yielding 56 valid unique participants. We then applied ten repeated subject-wise holdout splits and compared five compact baselines spanning Extra Trees and a multi-layer perceptron on asymmetry-aware spectral features and three convolutional networks on raw signals, including the EEG-specific EEGNet and ShallowConvNet architectures. Uncertainty was quantified through 95% bootstrap confidence intervals of the mean across repeats. The best model, an Extra Trees classifier using eight-channel spectral and asymmetry features, achieved a mean balanced accuracy of 93.5% with a 95% bootstrap confidence interval of 89.6% to 96.8% and a mean area under the receiver operating characteristic curve of 98.6% with a 95% bootstrap confidence interval of 96.2% to 100.0%. A connectivity ablation showed that inter-channel coherence was informative in isolation but did not improve performance when naively fused with spectral features. A feature-selection ablation did not show evidence that the 90-dimensional spectral representation was dominated by noisy or uninformative dimensions under this evaluation protocol. These results support compact, subject-wise evaluated EEG screening pipelines while highlighting the importance of rigorous leakage control.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repository

Its files are read in the Code ↔ Paper reader above, with 10 matches between paragraphs and lines of code.

ugail/EEG_Based_Depression_Screening

License: none: the authors keep all their rights

State: the link answers, verified on 28 September 2026

Evidence: files inventoried

Commit: 0a6888f3caaafcad451a447db807154eaa1f73d7, 15 May 2026

Languages: Jupyter (1)

Size: 9 files, 1 script

Software Heritage: not archived

Found in: “Data Availability Statement”

Holds: README, 1 notebook

Not found: license file, CITATION.cff, environment file, tests, continuous integration, documentation

Tools: Matplotlib (1 file), NumPy (1 file), pandas (1 file), PyTorch (1 file), scikit-learn (1 file), SciPy (1 file)

Availability: 1 check, the latest on 28 September 2026: the link answers

28 September 2026: the link answers

2 files

EEG_Depression_Analysis_Public_Notebook.ipynb, Jupyter, 1,297 lines, 10 matches
README.md, Text, 110 lines

At the source: github.com/ugail/EEG_Based_Depression_Screening

The paper's code and data availability statement is in the Data section.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
1 script, each with its path and the digest of its content;
10 matches between paragraphs of the paper and lines of the code (method lexical-v1);
neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

Datasets cited

figshare:4244171, at figshare; found in “Data Availability Statement”

Data Availability Statement

The EEG dataset analysed in this study is publicly available on figshare, DOI https://doi.org/10.6084/m9.figshare.4244171.v2. The code and data generated during the experimentation and analysis are publicly available at https://github.com/ugail/EEG_Based_Depression_Screening (accessed on 9 May 2026).

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 28 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 4 authors, 10 keywords, 11 MeSH terms, 1 funder, 29 references.

Cite

This paper

Ugail, H., Howard, N., Elmahmudi, A. A., & Mnasri, Z. (2026). Subject-Wise Depression Screening from Eight-Channel Resting-State EEG Using Asymmetry-Aware Spectral Features and Connectivity Ablation. Sensors (Basel, Switzerland), 26(10), 3065. https://doi.org/10.3390/s26103065

BibTeX

@article{ugail2026subject, author = {Ugail, Hassan and Howard, Newton and Elmahmudi, Ali Ahmed and Mnasri, Zied}, title = {{Subject-Wise Depression Screening from Eight-Channel Resting-State EEG Using Asymmetry-Aware Spectral Features and Connectivity Ablation}}, journal = {Sensors (Basel, Switzerland)}, year = {2026}, month = may, volume = {26}, number = {10}, pages = {3065}, publisher = {Multidisciplinary Digital Publishing Institute (MDPI)}, issn = {1424-8220}, doi = {10.3390/s26103065}, url = {https://doi.org/10.3390/s26103065}, pmid = {42197874}, pmcid = {PMC13210619} }

RIS

TY - JOUR AU - Ugail, Hassan AU - Howard, Newton AU - Elmahmudi, Ali Ahmed AU - Mnasri, Zied TI - Subject-Wise Depression Screening from Eight-Channel Resting-State EEG Using Asymmetry-Aware Spectral Features and Connectivity Ablation T2 - Sensors (Basel, Switzerland) J2 - Sensors (Basel) PY - 2026 DA - 2026/05/12 VL - 26 IS - 10 SP - 3065 SN - 1424-8220 PB - Multidisciplinary Digital Publishing Institute (MDPI) DO - 10.3390/s26103065 UR - https://doi.org/10.3390/s26103065 LA - en ER -

CSL-JSON

{ "id": "10.3390/s26103065", "type": "article-journal", "title": "Subject-Wise Depression Screening from Eight-Channel Resting-State EEG Using Asymmetry-Aware Spectral Features and Connectivity Ablation", "container-title": "Sensors (Basel, Switzerland)", "author": [ { "family": "Ugail", "given": "Hassan" }, { "family": "Howard", "given": "Newton" }, { "family": "Elmahmudi", "given": "Ali Ahmed" }, { "family": "Mnasri", "given": "Zied" } ], "container-title-short": "Sensors (Basel)", "volume": "26", "issue": "10", "page": "3065", "DOI": "10.3390/s26103065", "PMID": "42197874", "PMCID": "PMC13210619", "ISSN": "1424-8220", "publisher": "Multidisciplinary Digital Publishing Institute (MDPI)", "URL": "https://doi.org/10.3390/s26103065", "language": "en", "issued": { "date-parts": [ [ 2026, 5, 12 ] ] } }

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1007/s00422-026-01049-1 [code]: A three-component dynamical index of consciousness-related neural organisation.
Journal: Biological cybernetics
In common: scikit-learn, pandas, SciPy, 2 other tools, EEG, author Hassan Ugail
[2] doi:10.3389/fnins.2026.1874302 [code]: Architecture-data matching for EEG-EMG decoding: compact deep models match classical spectral decoders on the WAY-EEG-GAL grasp-and-lift dataset.
Journal: Frontiers in neuroscience
In common: PyTorch, scikit-learn, pandas, 3 other tools, other, EEG, 3 references
[3] doi:10.1371/journal.pone.0347671 [code]: RMETNet: A cross-subject motor imagery EEG signal classification model based on TSLANet and riemannian geometry features.
Journal: PloS one
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 3 references
[4] doi:10.1038/s41598-026-68186-2 [code]: NeuroStream: spectral-spatio-temporal deep learning for visual stimulus classification from EEG.
Journal: Scientific reports
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 3 references
[5] doi:10.1002/hbm.70528 [code]: Explainable AI Insights Into EEG Classification and Its Alignment to Neural Correlates.
Journal: Human brain mapping
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 2 references
[6] doi:10.1371/journal.pone.0351872 [code]: Decoding visual object recognition from EEG signals.
Journal: PloS one
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 2 references
[7] doi:10.1038/s41746-026-02778-0 [code]: Trust-gated synthetic EEG augmentation reduces performance drops when generalizing to new patients.
Journal: NPJ digital medicine
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 2 references
[8] doi:10.1038/s41398-026-04079-2 [code]: Neural Responses to Affective Sentences Reveal Signatures of Depression.
Journal: Translational psychiatry
In common: PyTorch, scikit-learn, pandas, 3 other tools, depression, EEG, 1 reference
[9] doi:10.3390/bios16070394: Hybrid Edge-Cloud Asymmetric Analytics for Portable Multimodal BCI Biosensors.
Journal: Biosensors
In common: other, EEG, 5 references
[10] doi:10.1186/s13634-026-01330-2 [code]: Leednet: a lightweight network for event detection in EEG signals.
Journal: Journal on advances in signal processing
In common: PyTorch, scikit-learn, pandas, 3 other tools, EEG, 2 references

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.