File size: 4,458 Bytes
ce4c4ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import numpy as np
import json
import gradio as gr
from transformers import AutoTokenizer
from captum.attr import IntegratedGradients
from torch_geometric.data import Data
from empath import Empath
import spacy

# -----------------------
# Devices
# -----------------------
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# -----------------------
# Load NLP
# -----------------------
try:
    nlp = spacy.load("en_core_web_sm")
except:
    import os
    os.system("python -m spacy download en_core_web_sm")
    nlp = spacy.load("en_core_web_sm")

empath = Empath()

# -----------------------
# Load Artifacts
# -----------------------
tokenizer = AutoTokenizer.from_pretrained("UFNLP/gatortron-base-2k")

with open("artifacts/union_trigrams.json") as f:
    TRIGRAM_LIST = json.load(f)

with open("artifacts/empath_cats.json") as f:
    EMPATH_CATS = json.load(f)

with open("artifacts/ip_op_trigram_sets.json") as f:
    sets = json.load(f)
    IP_SET = set(sets["ip"])
    OP_SET = set(sets["op"])

# -----------------------
# Model Definitions (same as training)
# -----------------------
from gatortron_gnn_captum import GatorTronEncoder, MetaGNN, GNNWrapper

ckpt = torch.load("artifacts/best_model.pt", map_location=DEVICE)

gatortron = GatorTronEncoder("UFNLP/gatortron-base-2k").to(DEVICE)
gatortron.load_state_dict(ckpt["gatortron"])
gatortron.eval()

gnn = MetaGNN(
    in_dim=ckpt["params"]["in_dim"],
    hidden_dim=ckpt["params"]["hidden_dim"],
    out_dim=2
).to(DEVICE)
gnn.load_state_dict(ckpt["gnn"])
gnn.eval()

# -----------------------
# Helpers
# -----------------------
def extract_trigrams(text):
    doc = nlp(text.lower())
    toks = [t.lemma_ for t in doc if t.is_alpha and not t.is_stop]
    return [" ".join(toks[i:i+3]) for i in range(len(toks)-2)]

def build_feature_vector(text):
    inp = tokenizer(
        text,
        truncation=True,
        padding="max_length",
        max_length=2000,
        return_tensors="pt"
    ).to(DEVICE)

    with torch.no_grad():
        gt = gatortron(inp["input_ids"], inp["attention_mask"]).cpu().numpy()[0]

    emp = empath.analyze(text, normalize=True)
    emp_vec = np.array([emp.get(c, 0.0) for c in EMPATH_CATS])

    trigs = extract_trigrams(text)
    tri_vec = np.array([trigs.count(t) for t in TRIGRAM_LIST])

    rsn = np.zeros(384)  # reasoning placeholder

    return np.concatenate([gt, emp_vec, tri_vec, rsn])

def explain(x_tensor):
    dummy_edge = torch.tensor([[0], [0]]).to(DEVICE)
    wrapper = GNNWrapper(gnn, dummy_edge)
    ig = IntegratedGradients(wrapper)

    attr = ig.attribute(
        x_tensor,
        baselines=torch.zeros_like(x_tensor),
        target=0,
        internal_batch_size=16
    )

    return attr.abs().cpu().numpy()[0]

# -----------------------
# Inference Function
# -----------------------
def predict(note):
    x = build_feature_vector(note)
    x_tensor = torch.tensor(x, dtype=torch.float32).unsqueeze(0).to(DEVICE)

    dummy_edge = torch.tensor([[0], [0]]).to(DEVICE)
    data = Data(x=x_tensor, edge_index=dummy_edge)

    with torch.no_grad():
        out = gnn(data)
        probs = torch.exp(out)[0].cpu().numpy()

    pred = "IP" if probs[0] > probs[1] else "OP"

    attr = explain(x_tensor)

    # ---- Empath ----
    emp_start = len(x) - (len(EMPATH_CATS) + len(TRIGRAM_LIST) + 384)
    emp_attr = attr[emp_start:emp_start+len(EMPATH_CATS)]

    top_empath = sorted(
        zip(EMPATH_CATS, emp_attr),
        key=lambda x: x[1],
        reverse=True
    )[:5]

    # ---- Trigrams ----
    tri_start = emp_start + len(EMPATH_CATS)
    tri_attr = attr[tri_start:tri_start+len(TRIGRAM_LIST)]

    top_trigrams = sorted(
        zip(TRIGRAM_LIST, tri_attr),
        key=lambda x: x[1],
        reverse=True
    )[:10]

    return (
        pred,
        float(probs[0]),
        float(probs[1]),
        top_empath,
        top_trigrams
    )

# -----------------------
# Gradio UI
# -----------------------
demo = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=12, label="Clinical Note"),
    outputs=[
        gr.Label(label="Prediction (IP / OP)"),
        gr.Number(label="IP Probability"),
        gr.Number(label="OP Probability"),
        gr.JSON(label="Top 5 Empath Categories"),
        gr.JSON(label="Top 10 Trigrams"),
    ],
    title="Clinical IP / OP Classifier with Explainability",
    description="GatorTron + GNN + Captum interpretability"
)

demo.launch()