HP_Opt.py

def get_time_series_smooth(ID,vitals):
    #pull the patients data
    vitals=vitals[vitals['patientunitstayid']==ID]
    vitals=vitals[['observationoffset','sao2']]
    vitals=vitals.fillna(method='pad')
    vitals=vitals.dropna()
    vital_times=np.array(vitals['observationoffset'])

    #smooth the necessary signals
    #from scipy.signal import savgol_filter
    sao2=vitals['sao2']
    sao2_transformed=1-np.exp((sao2-100)/10)
    sao2_smooth = np.concatenate((sao2_transformed[0:4],np.convolve(sao2_transformed, np.ones(5)/5,mode='valid')))

    
    #the patient time series
    #SaO2 only
    pt_series=np.transpose(np.vstack(sao2_smooth))

    return pt_series,vital_times

def time_data30(pt_series,vital_times):
    ys=[]
    xs=[]
    t=[]
    ts=2
    inputshape=0
    #for SaO2 model only
    for i in range(0,len(pt_series[0])-(ts+1)):
        y=pt_series[0][i+ts]
        t.append(vital_times[i+ts])
        #all
        x=pt_series[0][i:i+ts]
        ys.append(y)
        xs.append(x)
    ys=np.transpose(ys)
    n=0
    #downsampling
    d=6
    xs=xs[::d]
    ys=ys[::d]
    t=t[::d]
    return xs,ys,ts,t,n

def time_data5(pt_series,vital_times):
    ys=[]
    xs=[]
    t=[]
    ts=2
    inputshape=0
    #for SaO2 model only
    for i in range(0,len(pt_series[0])-(ts+1)):
        y=pt_series[0][i+ts]
        t.append(vital_times[i+ts])
        #all
        x=pt_series[0][i:i+ts]
        ys.append(y)
        xs.append(x)
    ys=np.transpose(ys)
    n=0
    
    return xs,ys,ts,t,n
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from sklearn.model_selection import GridSearchCV
from keras.wrappers.scikit_learn import KerasRegressor

def create_model(lr,Deep):
  lstm=tf.keras.Sequential()
  inputshape=1
  lstm.add(tf.keras.layers.BatchNormalization(input_shape=(ts,inputshape)))
  lstm.add(tf.keras.layers.LSTM(256,return_sequences=True,input_shape=(ts,inputshape)))
  if Deep==1:
    lstm.add(tf.keras.layers.Dropout(.1))
    lstm.add(tf.keras.layers.LSTM(128,return_sequences=True))
    #
    lstm.add(tf.keras.layers.Dropout(.1))
    lstm.add(tf.keras.layers.LSTM(64,return_sequences=True))
    #
    lstm.add(tf.keras.layers.Dropout(.1))
    lstm.add(tf.keras.layers.LSTM(32,return_sequences=True))

  lstm.add(tf.keras.layers.Dropout(.1))
  lstm.add(tf.keras.layers.LSTM(16,return_sequences=False))
  lstm.add(tf.keras.layers.Dense(units=1))
  opt = keras.optimizers.Adam(learning_rate=lr)
  lstm.compile(loss='mse',optimizer=opt)  
  return lstm


import pandas as pd
import numpy as np
vitals=pd.read_csv('vitals.csv')
TrainIDs=pd.read_csv('TrainEICU.csv',header=None)
TrainIDs=np.array(TrainIDs[0].astype(int))
VentTest=pd.read_csv('TestVent.csv',header=None)
VentTest=np.array(VentTest[0].astype(int))
NoVentTest=pd.read_csv('Test_NoVent.csv',header=None)
NoVentTest=np.array(NoVentTest[0].astype(int))

ID = TrainIDs[0]
pt_series,vital_times=get_time_series_smooth(ID,vitals)
#Have to change which time_data fxn to use depending on 30 min vs. 5 min
full_x_train,full_y_train,ts,t,n = time_data30(pt_series,vital_times)

True_TrainIDs=[]
for r in range(1,len(TrainIDs)):
    ID = TrainIDs[r]
    v=vitals[vitals['patientunitstayid']==ID]
    if len(v)>60:
        if ~(np.sum(v.isna()['sao2'])==len(v)):
            True_TrainIDs.append(ID)
            #need to pass ID, vents, vitals
            pt_series,vital_times=get_time_series_smooth(ID,vitals)
            xtrain,ytrain,ts,t,n = time_data30(pt_series,vital_times)
            full_x_train=np.vstack((full_x_train,xtrain))
            full_y_train=np.hstack((full_y_train,ytrain))

s=np.shape(full_x_train)[0]
xtrain=full_x_train.reshape(s,2,1)

model = KerasRegressor(build_fn=create_model, epochs=100,verbose=2)
# define the grid search parameters
lr=[.001,.01,.1]
Deep=[1,0]
param_grid = dict(lr=lr,Deep=Deep)
grid = GridSearchCV(estimator=model, param_grid=param_grid, cv=3,verbose=4)
grid_result = grid.fit(xtrain,full_y_train)

means = grid_result.cv_results_['mean_test_score']
stds = grid_result.cv_results_['std_test_score']
params = grid_result.cv_results_['params']

#Change names to run for 5 minute model
import joblib
joblib.dump(grid_result.best_params_, '30min_CV_bestparams.pkl')
joblib.dump(grid_result.best_score_, '30min_CV_bestscore.pkl')
joblib.dump(grid_result.cv_results_, '30min_CV_results.pkl')