-
Notifications
You must be signed in to change notification settings - Fork 2
/
shap_values.py
180 lines (156 loc) · 6.59 KB
/
shap_values.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
import argparse
import os
import numpy as np
import pandas as pd
import torch
import shap
from tqdm import tqdm
import matplotlib.pyplot as plt
from resnet import resnet34
from utils import prepare_input
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--data-dir', type=str, default='data/CPSC', help='Data directory')
parser.add_argument('--leads', type=str, default='all')
parser.add_argument('--seed', type=int, default=42, help='Seed to split data')
parser.add_argument('--use-gpu', default=False, action='store_true', help='Use GPU')
return parser.parse_args()
def plot_shap(ecg_data, sv_data, top_leads, patient_id, label):
# patient-level interpretation along with raw ECG data
leads = np.array(['I', 'II', 'III', 'aVR', 'aVL', 'aVF', 'V1', 'V2', 'V3', 'V4', 'V5', 'V6'])
nleads = len(top_leads)
if nleads == 0:
return
nsteps = 5000 # ecg_data.shape[1], visualize last 10 s since many patients' ECG are <=10 s
x = range(nsteps)
ecg_data = ecg_data[:, -nsteps:]
sv_data = sv_data[:, -nsteps:]
threshold = 0.001 # set threshold to highlight features with high shap values
fig, axs = plt.subplots(nleads, figsize=(9, nleads))
fig.suptitle(label)
for i, lead in enumerate(top_leads):
sv_upper = np.ma.masked_where(sv_data[lead] >= threshold, ecg_data[lead])
sv_lower = np.ma.masked_where(sv_data[lead] < threshold, ecg_data[lead])
if nleads == 1:
axe = axs
else:
axe = axs[i]
axe.plot(x, sv_upper, x, sv_lower)
axe.set_xticks([])
axe.set_yticks([])
axe.set_ylabel(leads[lead])
plt.savefig(f'shap/shap1-{patient_id}.png')
plt.close(fig)
def summary_plot(svs, y_scores):
leads = np.array(['I', 'II', 'III', 'aVR', 'aVL', 'aVF', 'V1', 'V2', 'V3', 'V4', 'V5', 'V6'])
svs2 = []
n = y_scores.shape[0]
for i in tqdm(range(n)):
label = np.argmax(y_scores[i])
sv_data = svs[label, i]
svs2.append(np.sum(sv_data, axis=1))
svs2 = np.vstack(svs2)
svs_data = np.mean(svs2, axis=0)
plt.plot(leads, svs_data)
plt.savefig('./shap/summary.png')
plt.clf()
def plot_shap2(svs, y_scores, cmap=plt.cm.Blues):
# population-level interpretation
leads = np.array(['I', 'II', 'III', 'aVR', 'aVL', 'aVF', 'V1', 'V2', 'V3', 'V4', 'V5', 'V6'])
n = y_scores.shape[0]
results = [[], [], [], [], [], [], [], [], []]
print(svs.shape)
for i in tqdm(range(n)):
label = np.argmax(y_scores[i])
results[label].append(svs[label, i])
ys = []
for label in range(y_scores.shape[1]):
result = np.array(results[label])
y = []
for i, _ in enumerate(leads):
y.append(result[:,i].sum())
y = np.array(y) / np.sum(y)
ys.append(y)
plt.plot(leads, y)
ys.append(np.array(ys).mean(axis=0))
ys = np.array(ys)
fig, axs = plt.subplots()
im = axs.imshow(ys, cmap=cmap)
axs.figure.colorbar(im, ax=axs)
fmt = '.2f'
xlabels = leads
ylabels = ['SNR', 'AF', 'IAVB', 'LBBB', 'RBBB', 'PAC', 'PVC', 'STD', 'STE'] + ['AVG']
axs.set_xticks(np.arange(len(xlabels)))
axs.set_yticks(np.arange(len(ylabels)))
axs.set_xticklabels(xlabels)
axs.set_yticklabels(ylabels)
thresh = ys.max() / 2
for i in range(ys.shape[0]):
for j in range(ys.shape[1]):
axs.text(j, i, format(ys[i, j], fmt),
ha='center', va='center',
color='white' if ys[i, j] > thresh else 'black')
np.set_printoptions(precision=2)
fig.tight_layout()
plt.savefig('./shap/shap2.png')
plt.clf()
if __name__ == '__main__':
args = parse_args()
data_dir = os.path.normpath(args.data_dir)
database = os.path.basename(data_dir)
args.model_path = f'models/resnet34_{database}_{args.leads}_{args.seed}.pth'
label_csv = os.path.join(data_dir, 'labels.csv')
reference_csv = os.path.join(data_dir, 'reference.csv')
lleads = np.array(['I', 'II', 'III', 'aVR', 'aVL', 'aVF', 'V1', 'V2', 'V3', 'V4', 'V5', 'V6'])
classes = np.array(['SNR', 'AF', 'IAVB', 'LBBB', 'RBBB', 'PAC', 'PVC', 'STD', 'STE'])
if args.use_gpu and torch.cuda.is_available():
device = torch.device('cuda:0')
else:
device = 'cpu'
if args.leads == 'all':
leads = 'all'
nleads = 12
else:
leads = args.leads.split(',')
nleads = len(leads)
model = resnet34(input_channels=nleads).to(device)
model.load_state_dict(torch.load(args.model_path, map_location=device))
model.eval()
background = 100
result_path = f'results/A{background * 2}.npy'
df_labels = pd.read_csv(label_csv)
df_reference = pd.read_csv(os.path.join(args.data_dir, 'reference.csv'))
df = pd.merge(df_labels, df_reference[['patient_id', 'age', 'sex', 'signal_len']], on='patient_id', how='left')
# df = df[df['signal_len'] >= 15000]
patient_ids = df['patient_id'].to_numpy()
to_explain = patient_ids[:background * 2]
background_patient_ids = df.head(background)['patient_id'].to_numpy()
background_inputs = [os.path.join(data_dir, patient_id) for patient_id in background_patient_ids]
background_inputs = torch.stack([torch.from_numpy(prepare_input(input)).float() for input in background_inputs]).to(device)
e = shap.GradientExplainer(model, background_inputs)
if not os.path.exists(result_path):
svs = []
y_scores = []
for patient_id in tqdm(to_explain):
input = os.path.join(data_dir, patient_id)
inputs = torch.stack([torch.from_numpy(prepare_input(input)).float()]).to(device)
y_scores.append(torch.sigmoid(model(inputs)).detach().cpu().numpy())
sv = np.array(e.shap_values(inputs)) # (n_classes, n_samples, n_leads, n_points)
svs.append(sv)
svs = np.concatenate(svs, axis=1)
y_scores = np.concatenate(y_scores, axis=0)
np.save(result_path, (svs, y_scores))
svs, y_scores = np.load(result_path, allow_pickle=True)
# summary_plot(svs, y_scores)
plot_shap2(svs, y_scores)
preds = []
top_leads_list = []
for i, patient_id in enumerate(to_explain):
ecg_data = prepare_input(os.path.join(data_dir, patient_id))
label_idx = np.argmax(y_scores[i])
sv_data = svs[label_idx, i]
sv_data_mean = np.mean(sv_data, axis=1)
top_leads = np.where(sv_data_mean > 1e-4)[0] # select top leads
preds.append(classes[label_idx])
print(patient_id, classes[label_idx], lleads[top_leads])
plot_shap(ecg_data, sv_data, top_leads, patient_id, classes[label_idx])