def step(self, action, reward_fun):
'''
action is a namedtuple with keys: basal, bolus
'''
CHO = 0.0
insulin = 0.0
BG = 0.0
CGM = 0.0
for _ in range(int(self.sample_time)):
# Compute moving average as the sample measurements
tmp_CHO, tmp_insulin, tmp_BG, tmp_CGM = self.mini_step(action)
CHO += tmp_CHO / self.sample_time
insulin += tmp_insulin / self.sample_time
BG += tmp_BG / self.sample_time
CGM += tmp_CGM / self.sample_time
# Compute risk index
horizon = 1
LBGI, HBGI, risk = risk_index([BG], horizon)
# Record current action
self.CHO_hist.append(CHO)
self.insulin_hist.append(insulin)
# Record next observation
self.time_hist.append(self.time)
self.BG_hist.append(BG)
self.CGM_hist.append(CGM)
self.risk_hist.append(risk)
self.LBGI_hist.append(LBGI)
self.HBGI_hist.append(HBGI)
# Compute reward, and decide whether game is over
window_size = int(60 / self.sample_time)
BG_last_hour = self.CGM_hist[-window_size:]
reward = reward_fun(bg_hist=self.BG_hist, cgm_hist=self.CGM_hist, insulin_hist=self.insulin_hist, risk_hist=self.risk_hist)
done = BG < 40 or BG > 350
obs = Observation(CGM=CGM)
return Step(
observation=obs,
reward=reward,
done=done,
sample_time=self.sample_time,
patient_name=self.patient.name,
meal=CHO,
patient_state=self.patient.state)
def _reset(self):
if self.perm_sample_time is None:
self.sample_time = self.sensor.sample_time
else:
self.sample_time = self.perm_sample_time
self.viewer = None
BG = self.patient.observation.Gsub
horizon = 1
LBGI, HBGI, risk = risk_index([BG], horizon)
CGM = self.sensor.measure(self.patient)
self.time_hist = [self.scenario.start_time]
self.BG_hist = [BG]
self.CGM_hist = [CGM]
self.risk_hist = [risk]
self.LBGI_hist = [LBGI]
self.HBGI_hist = [HBGI]
self.CHO_hist = [0]
self.insulin_hist = [0]
def step(self, action, reward_fun, cho):
'''
action is a namedtuple with keys: basal, bolus
'''
CHO = 0.0
insulin = 0.0
BG = 0.0
CGM = 0.0
if self.model is not None:
# Calculate CHO/insulin
for _ in range(int(self.sample_time)):
patient_action = self.scenario.get_action(self.time)
tmp_basal = self.pump.basal(action.basal)
tmp_bolus = self.pump.bolus(action.bolus)
tmp_insulin = tmp_basal + tmp_bolus
if cho is not None:
tmp_CHO = cho
else:
tmp_CHO = patient_action.meal
CHO += tmp_CHO / self.sample_time
insulin += tmp_insulin / self.sample_time
self.patient._t += 1 # copying mini-step of 1 minute
# Make state
state = np.concatenate([self.state, [CHO, insulin]])
norm_state = ((state-self.norm_params['mu'])/self.norm_params['std']).reshape(1, -1)
tensor_state = torch.from_numpy(norm_state).float().to(self.model_device)
# feed through model
with torch.no_grad():
next_state_tensor = self.model(tensor_state)
if self.model_device != 'cpu':
next_state_tensor = next_state_tensor.cpu()
next_state_norm = next_state_tensor.numpy().reshape(-1)
next_state = (next_state_norm*self.norm_params['std'][:13])+self.norm_params['mu'][:13]
self.state = next_state
# calculate BG and CGM
BG = self.state[12]/self.patient._params.Vg
self.patient._state[12] = self.state[12] # getting observation correct for CGM measurement
CGM = self.sensor.measure(self.patient)
else:
for _ in range(int(self.sample_time)):
# Compute moving average as the sample measurements
tmp_CHO, tmp_insulin, tmp_BG, tmp_CGM = self.mini_step(action, cho)
CHO += tmp_CHO / self.sample_time
insulin += tmp_insulin / self.sample_time
BG += tmp_BG / self.sample_time
CGM += tmp_CGM / self.sample_time
# Compute risk index
horizon = 1
LBGI, HBGI, risk = risk_index([BG], horizon)
magni_risk = magni_risk_index([BG])
# Record current action
self.CHO_hist.append(CHO)
self.insulin_hist.append(insulin)
# Record next observation
self.time_hist.append(self.time)
self.BG_hist.append(BG)
self.CGM_hist.append(CGM)
self.risk_hist.append(risk)
self.LBGI_hist.append(LBGI)
self.HBGI_hist.append(HBGI)
self.magni_risk_hist.append(magni_risk)
# Compute reward, and decide whether game is over
window_size = int(60 / self.sample_time)
BG_last_hour = self.CGM_hist[-window_size:]
reward = reward_fun(bg_hist=self.BG_hist, cgm_hist=self.CGM_hist, insulin_hist=self.insulin_hist,
risk_hist=self.risk_hist)
done = BG < 40 or BG > 350
obs = Observation(CGM=CGM)
return Step(
observation=obs,
reward=reward,
done=done,
sample_time=self.sample_time,
patient_name=self.patient.name,
meal=CHO,
patient_state=self.patient.state)
def risk_bg(bg_hist, **kwargs):
return -risk_index([bg_hist[-1]], 1)[-1]
def risk_diff_bg(bg_hist, **kwargs):
if len(bg_hist) < 2:
return 0
_, _, risk_current = risk_index([bg_hist[-1]], 1)
_, _, risk_prev = risk_index([bg_hist[-2]], 1)
return risk_prev - risk_current