add single channel decoding

decode_eyes_open_closed.py

@@ -6,6 +6,7 @@
 import numpy as np
 from sklearn import metrics, model_selection, linear_model
 from joblib import Parallel, delayed
+from catboost import CatBoostClassifier
 
 PATH_FEATURES = r"C:\Users\ICN_admin\OneDrive - Charité - Universitätsmedizin Berlin\Dokumente\Decoding toolbox\EyesOpenBeijing\2708\features_all.csv"
 def balance_classes(df, target_column):
@@ -25,6 +26,7 @@ def run_cv(df):
     model = linear_model.LinearRegression()
     cm_list = []
     ba_list = []
+    coef_list = []
     for train_, test_ in cv.split(df, df["label_enc"]):
         # integer encode the label column
         train_df = df.iloc[train_]
@@ -34,8 +36,6 @@ def run_cv(df):
         y_train = train_df_balanced["label_enc"]
 
         # ensure that the same number of samples are taken for training from each class
-
-
         X_test = df.iloc[test_, :]
         y_test = df.iloc[test_, :]["label_enc"]
         X_test = X_test.drop(columns=["label_enc"])
@@ -44,9 +44,9 @@ def run_cv(df):
         #cols_use = [c for c in X_train.columns if all([f in c for f in features])]
         #X_train = X_train[cols_use]
         #X_test = X_test[cols_use]
-
 
         model = linear_model.LogisticRegression(class_weight="balanced")
+        model = CatBoostClassifier(verbose=0)
         model.fit(X_train, y_train)
         pred = model.predict(X_test)
 
@@ -60,7 +60,10 @@ def run_cv(df):
 
         cm_list.append(cm)
         ba_list.append(metrics.balanced_accuracy_score(y_test, pred))
-
+        if type(model) == linear_model.LogisticRegression:
+            coef_list.append(model.coef_)
+        else:
+            coef_list.append(model.get_feature_importance())
         print(metrics.balanced_accuracy_score(y_test, pred))
 
     # plot the mean confusion matrix
@@ -71,31 +74,43 @@ def run_cv(df):
     #plt.show(block=True)
 
     ba_mean = sum(ba_list) / len(ba_list)
-    return ba_mean, cm_mean
+    return ba_mean, cm_mean, np.mean(coef_list, axis=0)
 
-def compute_modality(mod):
+def compute_modality(sub, mod):
+
+
+
+    #for mod in modality_:
     l_df = []
     if mod != "all":
         df_mod = df_all[[c for c in df_all.columns if mod in c] + ["label_enc"] + ["sub"] + ["disease"]].copy()
     else:
         df_mod = df_all.copy()
 
-    for sub in subs:
-        df_sub = df_mod[df_mod["sub"] == sub]
-        disease = df_sub["disease"].unique()[0]
-        for loc in locs:
-            cols_loc = [c for c in df_sub.columns if loc in c] + ["label_enc"]
-            if len(cols_loc) == 0:
-                continue
-            df_sub_loc = df_sub[cols_loc].copy()
-            # select only columns that have non NaN values
-            df_sub_loc = df_sub_loc.dropna(axis=1)
-            if len(df_sub_loc.columns) == 1:
-                continue
-
-            ba_mean, cm_mean = run_cv(df_sub_loc.copy())
-
-            l_df.append({"sub": sub, "loc": loc, "mod": mod, "disease": disease, "ba": ba_mean})
+    df_sub = df_mod[df_mod["sub"] == sub]
+    disease = df_sub["disease"].unique()[0]
+    for loc in locs:
+        print(loc)
+        cols_loc = [c for c in df_sub.columns if loc in c] + ["label_enc"]
+        if len(cols_loc) == 0:
+            continue
+        df_sub_loc = df_sub[cols_loc].copy()
+        # select only columns that have non NaN values
+        df_sub_loc = df_sub_loc.dropna(axis=1)
+        if len(df_sub_loc.columns) == 1:
+            continue
+
+        chs_ = np.unique([c.split("_")[0] for c in df_sub_loc.columns if "label" not in c])
+        for ch in chs_:
+            print(ch)
+            df_sub_ch = df_sub_loc[[c for c in df_sub_loc.columns if ch in c or "label" in c]].copy()
+            f_bands = [c.split("_")[2] for c in df_sub_ch.columns if "label" not in c]
+            ba_mean, cm_mean, coef_mean = run_cv(df_sub_ch.copy())
+
+            dict_out = {"sub": sub, "loc": loc, "ch" : str(ch), "mod": mod, "dout": disease, "ba": float(ba_mean)}
+            for i, coef in enumerate(coef_mean.T):
+                dict_out[f"coef_{f_bands[i]}"] = float(coef)
+            l_df.append(dict_out)
     return l_df
 
 if __name__ == "__main__":
@@ -109,36 +124,41 @@ def compute_modality(mod):
 
     locs = ["STN", "GPI", "ECOG", "EEG"]
     modality_ = ["theta", "alpha", "low beta", "high beta", "low gamma", "high gamma", "HFA", "fft", "fooof", "all"]
+    modality_ = ["fft", "alpha", ]
 
     subs = df_all["sub"].unique()
     diseases = df_all["disease"].unique()
 
-    #compute_modality(modality_[0])
+    #compute_modality(subs[0], modality_[0])
 
-    l_df_ = Parallel(n_jobs=len(modality_))(delayed(compute_modality)(mod) for mod in modality_)
+    PATH_BASE = r"C:\Users\ICN_admin\OneDrive - Charité - Universitätsmedizin Berlin\Dokumente\Decoding toolbox\EyesOpenBeijing\2708"
+    for mod in modality_:
+        l_df_ = Parallel(n_jobs=len(modality_))(delayed(compute_modality)(sub, mod) for sub in subs)
 
+        df_per = pd.DataFrame(list(np.concat(l_df_)))
+
+        df_per.to_csv(os.path.join(PATH_BASE, f"out_per_loc_mod_{mod}_CB.csv"), index=False)
 
 
 
 
-    df_per = pd.DataFrame(list(np.concat(l_df_)))
-    df_per.to_csv(r"C:\Users\ICN_admin\OneDrive - Charité - Universitätsmedizin Berlin\Dokumente\Decoding toolbox\EyesOpenBeijing\2708\out_per_loc_mod.csv", index=False)
 
-    plt.figure()
-    sns.boxplot(data=df_per, x="loc", y="ba", hue="mod", palette="viridis")
-    #sns.swarmplot(data=df_per, x="loc", y="ba", color=".25", hue="mod")
-    plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
-    plt.tight_layout()
-    plt.savefig("ba_")
-    plt.show(block=True)
 
-    plt.figure()
-    sns.boxplot(data=df_per.query("mod == 'all'"), x="disease", y="ba", hue="mod", palette="viridis")
-    sns.swarmplot(data=df_per.query("mod == 'all'"), x="disease", y="ba", color=".25", hue="mod")
-    plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
-    plt.tight_layout()
-    plt.savefig("ba_")
-    plt.show(block=True)
+    # plt.figure()
+    # sns.boxplot(data=df_per, x="loc", y="ba", hue="mod", palette="viridis")
+    # #sns.swarmplot(data=df_per, x="loc", y="ba", color=".25", hue="mod")
+    # plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
+    # plt.tight_layout()
+    # plt.savefig("ba_")
+    # plt.show(block=True)
+
+    # plt.figure()
+    # sns.boxplot(data=df_per.query("mod == 'all'"), x="disease", y="ba", hue="mod", palette="viridis")
+    # sns.swarmplot(data=df_per.query("mod == 'all'"), x="disease", y="ba", color=".25", hue="mod")
+    # plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
+    # plt.tight_layout()
+    # plt.savefig("ba_")
+    # plt.show(block=True)
 
     # for loc in locs:
     #     features = ["gamma",]

pyproject.toml

@@ -41,7 +41,6 @@ dependencies = [
     "numpy >= 1.21.2",
     "pandas >= 2.0.0",
     "scikit-image",
-    "scikit-learn >= 0.24.2",
     "scikit-optimize",
     "scipy >= 1.7.1",
     "seaborn >= 0.11",
@@ -57,6 +56,7 @@ dependencies = [
     "pydantic>=2.8.2",
     "mne-qt-browser>=0.6.3",
     "specparam>=2.0.0rc2",
+    "catboost>=1.2.7",
 ]
 
 [project.optional-dependencies]

read_res_decoding_single_ch.py

@@ -0,0 +1,34 @@
+import pandas as pd
+import os
+
+from matplotlib import pyplot as plt
+import seaborn as sns
+
+PATH_ = r"C:\Users\ICN_admin\OneDrive - Charité - Universitätsmedizin Berlin\Dokumente\Decoding toolbox\EyesOpenBeijing\2708"
+
+mods = ["alpha", "fft"]
+
+l_ = []
+for mod in mods:
+    df = pd.read_csv(os.path.join(PATH_, f"out_per_loc_mod_{mod}.csv"))
+    df["mod"] = mod
+    l_.append(df)
+
+df_all = pd.concat(l_)
+
+plt.figure()
+sns.boxplot(data=df_all, x="loc", y="ba", hue="mod", palette="viridis")
+plt.tight_layout()
+plt.show(block=True)
+
+df = pd.read_csv(os.path.join(PATH_, f"out_per_loc_mod_fft.csv"))
+# melt the dataframe that all columns with coef_ become a column
+df_melt = df.melt(id_vars=["ba", "loc", "sub", "dout"], value_vars=[c for c in df.columns if "coef_" in c])
+
+plt.figure()
+# melt the dataframe that coef_ becomes a column
+sns.boxplot(data=df_melt, x="variable", y="value", palette="viridis")
+plt.xticks(rotation=90)
+plt.ylabel("Coef")
+plt.tight_layout()
+plt.show(block=True)