IMH-XGBoost / app.py
Limour-blog's picture
Upload 2 files
4bb4a0c verified
# conda create -n IMH-XGBoost conda-forge::huggingface_hub
# pip install -r requirements.txt -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
import os
# 获取模型
if not os.path.exists('xgb.baseline.model.json'):
from huggingface_hub import login, snapshot_download
login(token=os.environ.get("HF_TOKEN"))
snapshot_download(repo_id='Limour-blog/IMH-XGBoost', local_dir=r'.', allow_patterns='xgb.baseline.model.json')
import xgboost as xgb
import numpy as np
clf = xgb.XGBClassifier(enable_categorical=True)
clf.load_model(r"xgb.baseline.model.json")
def limit(_value, _min, _max):
return min(max(_value, _min), _max)
def args2Array(
BSA=1.824,
CTNT=4.715, # _0
CK_MB=200.5, # _0
CRP=18.01, # _1
PD_DIMER=1.047,
NT_PROBNP=883.6, # _3
ARRHYTHMIA=0,
APOE=36.76,
MHR=0.8378
):
BSA = limit(BSA, 1.401, 2.231)
BSA = (BSA - 1.824) / 0.1654
CTNT = limit(CTNT, -9.566, 19.58)
CTNT = (CTNT - 4.715) / 3.877
CK_MB = limit(CK_MB, -213, 571)
CK_MB = (CK_MB - 200.5) / 154.3
CRP = limit(CRP, -25.04, 55.86)
CRP = (CRP - 18.01) / 17.53
PD_DIMER = limit(PD_DIMER, -1.131, 2.959)
PD_DIMER = (PD_DIMER - 1.047) / 0.8045
NT_PROBNP = limit(NT_PROBNP, -610.1, 2106)
NT_PROBNP = (NT_PROBNP - 883.6) / 625.8
APOE = limit(APOE, 3.625, 68.62)
APOE = (APOE - 36.76) / 13.85
MHR = limit(MHR, -0.06439, 1.683)
MHR = (MHR - 0.8378) / 0.3103
return np.array([[BSA, CTNT, CK_MB,
CRP, PD_DIMER, NT_PROBNP,
ARRHYTHMIA, APOE, MHR]])
def predict(_array):
return float(clf.predict_proba(_array)[0,1])
# 测试模型预测阳性正确
assert predict(args2Array(
BSA=1.99,
CTNT=10, # _0
CK_MB=374, # _0
CRP=14.4, # _1
PD_DIMER=0.88,
NT_PROBNP=463.7, # _3
ARRHYTHMIA=0,
APOE=37,
MHR=0.8378
)) >= 0.72
# 测试模型预测阴性正确
assert predict(args2Array(
BSA=1.51,
CTNT=1.53, # _0
CK_MB=95, # _0
CRP=4.9, # _1
PD_DIMER=1.4,
NT_PROBNP=519.2, # _3
ARRHYTHMIA=0,
APOE=36.76,
MHR=0.5581
)) < 0.72
import gradio as gr
# ========== 完整版的模型 ==========
with gr.Blocks() as complete_model:
with gr.Row():
g_BSA = gr.Number(label="BSA", scale=1, value=1.824,
info="患者的体表面积, 缺失请保持默认值",
interactive=True)
g_ARRHYTHMIA = gr.Checkbox(label="ARRHYTHMIA", scale=1, value=False,
info="患者是否发生恶性心律失常或传导阻滞, 缺失请保持默认值",
interactive=True)
g_PD_DIMER = gr.Number(label="PD_DIMER", scale=1, value=1.047,
info="PCI术后D-二聚体峰值, 缺失请保持默认值",
interactive=True)
with gr.Row():
g_CTNT = gr.Number(label="CTNT", scale=1, value=4.715,
info="PCI术后即刻的CTNT值, 缺失请保持默认值",
interactive=True)
g_CK_MB = gr.Number(label="CK_MB", scale=1, value=200.5,
info="PCI术后即刻的CK_MB值, 缺失请保持默认值",
interactive=True)
g_NT_PROBNP = gr.Number(label="NT_PROBNP", scale=1, value=883.6,
info="PCI术后36小时的NT_PROBNP值, 缺失请保持默认值",
interactive=True)
with gr.Row():
g_CRP = gr.Number(label="CRP", scale=1, value=18.01,
info="PCI术后24小时的CRP值, 缺失请保持默认值",
interactive=True)
g_APOE = gr.Number(label="APOE", scale=1, value=36.76,
info="患者血脂APOE值, 缺失请保持默认值",
interactive=True)
g_MHR = gr.Number(label="MHR", scale=1, value=0.8378,
info="单核细胞与高密度脂蛋白胆固醇比值, 缺失请保持默认值",
interactive=True)
with gr.Row():
g_output1 = gr.Number(label="XGB.predict_proba", scale=1, interactive=False, info="cutoff值为0.72")
g_output2 = gr.Textbox(label="结论", scale=1, interactive=False, info="预测患者IMH为阳性或阴性")
g_calc = gr.Button("计算", variant="primary", size='lg')
def btn_calc(
BSA, CTNT, CK_MB,
CRP, PD_DIMER, NT_PROBNP,
ARRHYTHMIA, APOE, MHR
):
res1 = predict(args2Array(
BSA=BSA,
CTNT=CTNT, # _0
CK_MB=CK_MB, # _0
CRP=CRP, # _1
PD_DIMER=PD_DIMER,
NT_PROBNP=NT_PROBNP, # _3
ARRHYTHMIA = (1 if ARRHYTHMIA else 0),
APOE=APOE,
MHR=MHR
))
if res1 >= 0.72:
res2 = '阳性'
else:
res2 = '阴性'
return round(res1, 4), res2
g_calc.click(
fn = btn_calc,
inputs=[g_BSA, g_CTNT, g_CK_MB,
g_CRP, g_PD_DIMER, g_NT_PROBNP,
g_ARRHYTHMIA, g_APOE, g_MHR],
outputs=[g_output1, g_output2]
)
# ========== 开始运行 ==========
demo = gr.TabbedInterface([complete_model],
["complete_model"])
gr.close_all()
demo.queue(api_open=False, max_size=1).launch(
server_name = "0.0.0.0",
share=False, show_error=True, show_api=False)