def evaluate_solution(still_x, still_y, still_z, still_n, calibration_parameters):
""" Calculates the RMSE of the input XYZ signal if calibrated according to input calibration parameters"""
# Temporarily adjust the channels of still data, which has collapsed x,y,z values
do_calibration(still_x, still_y, still_z, calibration_parameters)
# Get the VM of the calibrated channel
vm = channel_inference.infer_vector_magnitude(still_x, still_y, still_z)
# se = sum error
se = 0.0
for vm_val,n in zip(vm.data, still_n.data):
se += (abs(1.0 - vm_val)**2)*n
rmse = math.sqrt(se / len(vm.data))
# Undo the temporary calibration
undo_calibration(still_x, still_y, still_z, calibration_parameters)
return rmse
from datetime import datetime, date, time, timedelta
from pampro import data_loading, Time_Series, Channel, channel_inference, triaxial_calibration
# Change filenames as appropriate
# Read the data - yields 1 channel per axis
ts, header = data_loading.load("/pa/data/BIOBANK/example.cwa", "Axivity")
x, y, z = ts.get_channels(["X", "Y", "Z"])
# Autocalibrate the raw acceleration data
x, y, z, calibration_diagnostics = triaxial_calibration.calibrate(x, y, z)
# Infer some sample level information - Vector Magnitude (VM), Euclidean Norm Minus One (ENMO)
vm = channel_inference.infer_vector_magnitude(x, y, z)
enmo = channel_inference.infer_enmo(vm)
# Create a time series object and add channels to it
ts.add_channels([vm, enmo])
# Uncomment this line to write the raw data as CSV
#ts.write_channels_to_file("C:/Data/3.csv")
# Request some interesting statistics - mean of ENMO
stats = {"ENMO":[("generic", ["mean"])]}
# Get the above statistics on an hourly level - returned as channels
hourly_results = ts.piecewise_statistics(timedelta(hours=1), statistics=stats)
# Write the hourly analysis to a file
from datetime import datetime, date, time, timedelta
from pampro import Time_Series, Channel, channel_inference, triaxial_calibration
# Change the filenames as appropriate
# Load sample activPAL data
x, y, z = Channel.load_channels(
"/pa/data/STVS/_data/activpal_data/714952C-AP1335893 18Nov13 10-00am for 7d 23h 14m.datx",
"activPAL")
# Autocalibrate the raw acceleration data
x, y, z, (cal_params), (results), (misc) = triaxial_calibration.calibrate(
x, y, z)
# Infer some sample level info from the three channels - VM, ENMO, Pitch & Roll
vm = channel_inference.infer_vector_magnitude(x, y, z)
enmo = channel_inference.infer_enmo(vm)
pitch, roll = channel_inference.infer_pitch_roll(x, y, z)
# Create a time series object and add all signals to it
ts = Time_Series.Time_Series("activPAL")
ts.add_channels([x, y, z, vm, enmo, pitch, roll])
# Request some stats about the time series
# In this case: mean ENMO, pitch and roll, and 10 degree cutpoints of pitch and roll
angle_levels = [[-90, -80], [-80, -70], [-70, -60], [-60, -50], [-50, -40],
[-40, -30], [-30, -20], [-20, -10], [-10, 0], [0, 10],
[10, 20], [20, 30], [30, 40], [40, 50], [50, 60], [60, 70],
[70, 80], [80, 90]]
stat_dict = {
"Pitch": angle_levels + ["mean"],
def qc_analysis(job_details):
id_num = str(job_details["pid"])
filename = job_details["filename"]
filename_short = os.path.basename(filename).split('.')[0]
battery_max = 0
if filetype == "GeneActiv":
battery_max = GA_battery_max
elif filetype == "Axivity":
battery_max = AX_battery_max
# Load the data from the hdf5 file
ts, header = data_loading.fast_load(filename, filetype)
header["QC_filename"] = os.path.basename(filename)
x, y, z, battery, temperature = ts.get_channels(["X", "Y", "Z", "Battery", "Temperature"])
# create a channel of battery percentage, based on the assumed battery maximum value
battery_pct = Channel.Channel.clone(battery)
battery_pct.data = (battery.data / battery_max) * 100
channels = [x, y, z, battery, temperature, battery_pct]
anomalies = diagnostics.diagnose_fix_anomalies(channels, discrepancy_threshold=2)
# create dictionary of anomalies types
anomalies_dict = dict()
# check whether any anomalies have been found:
if len(anomalies) > 0:
anomalies_file = os.path.join(anomalies_folder, "{}_anomalies.csv".format(filename_short))
df = pd.DataFrame(anomalies)
for type in anomaly_types:
anomalies_dict["QC_anomaly_{}".format(type)] = (df.anomaly_type.values == type).sum()
df = df.set_index("anomaly_type")
# print record of anomalies to anomalies_file
df.to_csv(anomalies_file)
else:
for type in anomaly_types:
anomalies_dict["QC_anomaly_{}".format(type)] = 0
# check for axis anomalies
axes_dict = diagnostics.diagnose_axes(x, y, z, noise_cutoff_mg=13)
axis_anomaly = False
for key, val in axes_dict.items():
anomalies_dict["QC_{}".format(key)] = val
if key.endswith("max"):
if val > axis_max:
axis_anomaly = True
elif key.endswith("min"):
if val < axis_min:
axis_anomaly = True
# create a "check battery" flag:
check_battery = False
# calculate first and last battery percentages
first_battery_pct = round((battery_pct.data[1]),2)
last_battery_pct = round((battery_pct.data[-1]),2)
header["QC_first_battery_pct"] = first_battery_pct
header["QC_last_battery_pct"] = last_battery_pct
# calculate lowest battery percentage
# check if battery.pct has a missing_value, exclude those values if they exist
if battery_pct.missing_value == "None":
lowest_battery_pct = min(battery_pct.data)
else:
test_array = np.delete(battery_pct.data, np.where(battery_pct.data == battery_pct.missing_value))
lowest_battery_pct = min(test_array)
header["QC_lowest_battery_pct"] = round(lowest_battery_pct,2)
header["QC_lowest_battery_threshold"] = battery_minimum
# find the maximum battery discharge in any 24hr period:
max_discharge = battery_pct.channel_max_decrease(time_period=timedelta(hours=discharge_hours))
header["QC_max_discharge"] = round(max_discharge, 2)
header["QC_discharge_time_period"] = "{} hours".format(discharge_hours)
header["QC_discharge_threshold"] = discharge_pct
# change flag if lowest battery percentage dips below battery_minimum at any point
# OR maximum discharge greater than discharge_pct over time period "hours = discharge_hours"
if lowest_battery_pct < battery_minimum or max_discharge > discharge_pct:
check_battery = True
header["QC_check_battery"] = str(check_battery)
header["QC_axis_anomaly"] = str(axis_anomaly)
# Calculate the time frame to use
start = time_utilities.start_of_day(x.timeframe[0])
end = time_utilities.end_of_day(x.timeframe[-1])
tp = (start, end)
results_ts = Time_Series.Time_Series("")
# Derive some signal features
vm = channel_inference.infer_vector_magnitude(x, y, z)
enmo = channel_inference.infer_enmo(vm)
enmo.minimum = 0
enmo.maximum = enmo_max
# Infer nonwear
nonwear_bouts = channel_inference.infer_nonwear_for_qc(x, y, z, noise_cutoff_mg=noise_cutoff_mg)
# Use nonwear bouts to calculate wear bouts
wear_bouts = Bout.time_period_minus_bouts(enmo.timeframe, nonwear_bouts)
# Use wear bouts to calculate the amount of wear time in the file in hours, save to meta data
total_wear = Bout.total_time(wear_bouts)
total_seconds_wear = total_wear.total_seconds()
total_hours_wear = round(total_seconds_wear/3600)
header["QC_total_hours_wear"] = total_hours_wear
# Split the enmo channel into lists of bouts for each quadrant:
''' quadrant_0 = 00:00 -> 06: 00
quadrant_1 = 06:00 -> 12: 00
quadrant_2 = 12:00 -> 18: 00
quadrant_3 = 18:00 -> 00: 00 '''
q_0, q_1, q_2, q_3 = channel_inference.create_quadrant_bouts(enmo)
# calculate the intersection of each set of bouts with wear_bouts, then calculate the wear time in each quadrant.
sum_quadrant_wear = 0
for quadrant, name1, name2 in ([q_0, "QC_hours_wear_quadrant_0", "QC_pct_wear_quadrant_0"],
[q_1, "QC_hours_wear_quadrant_1", "QC_pct_wear_quadrant_1"],
[q_2, "QC_hours_wear_quadrant_2", "QC_pct_wear_quadrant_2"],
[q_3, "QC_hours_wear_quadrant_3", "QC_pct_wear_quadrant_3"]):
quadrant_wear = Bout.bout_list_intersection(quadrant, wear_bouts)
seconds_wear = Bout.total_time(quadrant_wear).total_seconds()
hours_wear = round(seconds_wear / 3600)
header[name1] = hours_wear
header[name2] = round(((hours_wear / total_hours_wear) * 100), 2)
for bout in nonwear_bouts:
# Show non-wear bouts in purple
bout.draw_properties = {'lw': 0, 'alpha': 0.75, 'facecolor': '#764af9'}
for channel, channel_name in zip([enmo, battery_pct],["ENMO", "Battery_percentage"]):
channel.name = channel_name
results_ts.add_channel(channel)
if PLOT == "YES":
# Plot statistics as subplots in one plot file per data file
results_ts["ENMO"].add_annotations(nonwear_bouts)
results_ts.draw_qc(plotting_df, file_target=os.path.join(charts_folder,"{}_plots.png".format(filename_short)))
header["QC_script"] = version
# file of metadata from qc process
qc_output = os.path.join(results_folder, "qc_meta_{}.csv".format(filename_short))
# check if qc_output already exists...
if os.path.isfile(qc_output):
os.remove(qc_output)
metadata = {**header, **anomalies_dict}
# write metadata to file
pampro_utilities.dict_write(qc_output, id_num, metadata)
for c in ts:
del c.data
del c.timestamps
del c.indices
del c.cached_indices
def process_file(job_details):
id_num = str(job_details["pid"])
filename = job_details["filename"]
filename_short = os.path.basename(filename).split('.')[0]
meta = os.path.join(results_folder,
"metadata_{}.csv".format(filename_short))
# check if analysis_meta already exists...
if os.path.isfile(meta):
os.remove(meta)
battery_max = 0
if monitor_type == "GeneActiv":
battery_max = GA_battery_max
elif monitor_type == "Axivity":
battery_max = AX_battery_max
epochs = [timedelta(minutes=n) for n in epoch_minutes]
# Use 'epochs_minutes' variable to create the corresponding names to the epochs defined
names = []
plots_list = []
for n in epoch_minutes:
name = ""
if n % 60 == 0: # If the epoch is a multiple of 60, it will be named in hours, e.g. '1h'
name = "{}h".format(int(n / 60))
elif n % 60 != 0: # If the epoch is NOT a multiple of 60, it will be named in seconds, e.g. '15m'
name = "{}m".format(n)
names.append(name)
if n in epoch_plot:
plots_list.append(name)
# fast-load the data to identify any anomalies:
qc_ts, qc_header = data_loading.fast_load(filename, monitor_type)
qc_channels = qc_ts.get_channels(["X", "Y", "Z"])
anomalies = diagnostics.diagnose_fix_anomalies(qc_channels,
discrepancy_threshold=2)
# Load the data
ts, header = data_loading.load(filename, monitor_type, compress=False)
header["processed_file"] = os.path.basename(filename)
# some monitors have manufacturers parameters applied to them, let's preserve these but rename them:
var_list = [
"x_gain", "x_offset", "y_gain", "y_offset", "z_gain", "z_offset",
"calibration_date"
]
for var in var_list:
if var in header.keys():
header[("manufacturers_%s" % var)] = header[var]
header.pop(var)
x, y, z, battery, temperature, integrity = ts.get_channels(
["X", "Y", "Z", "Battery", "Temperature", "Integrity"])
initial_channels = [x, y, z, battery, temperature, integrity]
# create dictionary of anomalies total and types
anomalies_dict = {"QC_anomalies_total": len(anomalies)}
# check whether any anomalies have been found:
if len(anomalies) > 0:
anomalies_file = os.path.join(
results_folder, "{}_anomalies.csv".format(filename_short))
df = pd.DataFrame(anomalies)
for type in anomaly_types:
anomalies_dict["QC_anomaly_{}".format(type)] = (
df.anomaly_type.values == type).sum()
df = df.set_index("anomaly_type")
# print record of anomalies to anomalies_file
df.to_csv(anomalies_file)
# if anomalies have been found, fix these anomalies
channels = diagnostics.fix_anomalies(anomalies, initial_channels)
else:
for type in anomaly_types:
anomalies_dict["QC_anomaly_{}".format(type)] = 0
# if no anomalies
channels = initial_channels
first_channel = channels[0]
# Convert timestamps to offsets from the first timestamp
start, offsets = Channel.timestamps_to_offsets(first_channel.timestamps)
# As timestamps are sparse, expand them to 1 per observation
offsets = Channel.interpolate_offsets(offsets, len(first_channel.data))
# For each channel, convert to offset timestamps
for c in channels:
c.start = start
c.set_contents(c.data, offsets, timestamp_policy="offset")
# find approximate first and last battery percentage values
first_battery_pct = round((battery.data[1] / battery_max) * 100, 2)
last_battery_pct = round((battery.data[-1] / battery_max) * 100, 2)
# Calculate the time frame to use
start = time_utilities.start_of_day(x.timeframe[0])
end = time_utilities.end_of_day(x.timeframe[-1])
tp = (start, end)
# if the sampling frequency is greater than 40Hz
if x.frequency > 40:
# apply a low pass filter
x = pampro_fourier.low_pass_filter(x,
20,
frequency=x.frequency,
order=4)
x.name = "X" # because LPF^ changes the name, we want to override that
y = pampro_fourier.low_pass_filter(y,
20,
frequency=y.frequency,
order=4)
y.name = "Y"
z = pampro_fourier.low_pass_filter(z,
20,
frequency=z.frequency,
order=4)
z.name = "Z"
# find any bouts where data is "missing" BEFORE calibration
missing_bouts = []
if -111 in x.data:
# extract the bouts of the data channels where the data == -111 (the missing value)
missing = x.bouts(-111, -111)
# add a buffer of 2 minutes (120 seconds) to the beginning and end of each bout
for item in missing:
bout_start = max(item.start_timestamp - timedelta(seconds=120),
x.timeframe[0])
bout_end = min(item.end_timestamp + timedelta(seconds=120),
x.timeframe[1])
new_bout = Bout.Bout(start_timestamp=bout_start,
end_timestamp=bout_end)
missing_bouts.append(new_bout)
else:
pass
x.delete_windows(missing_bouts)
y.delete_windows(missing_bouts)
z.delete_windows(missing_bouts)
integrity.fill_windows(missing_bouts, fill_value=1)
################ CALIBRATION #######################
# extract still bouts
calibration_ts, calibration_header = triaxial_calibration.calibrate_stepone(
x, y, z, noise_cutoff_mg=noise_cutoff_mg)
# Calibrate the acceleration to local gravity
cal_diagnostics = triaxial_calibration.calibrate_steptwo(
calibration_ts, calibration_header, calibration_statistics=False)
# calibrate data
triaxial_calibration.do_calibration(x,
y,
z,
temperature=None,
cp=cal_diagnostics)
x.delete_windows(missing_bouts)
y.delete_windows(missing_bouts)
z.delete_windows(missing_bouts)
temperature.delete_windows(missing_bouts)
battery.delete_windows(missing_bouts)
# Derive some signal features
vm = channel_inference.infer_vector_magnitude(x, y, z)
vm.delete_windows(missing_bouts)
if "HPFVM" in stats:
vm_hpf = channel_inference.infer_vm_hpf(vm)
else:
vm_hpf = None
if "ENMO" in stats:
enmo = channel_inference.infer_enmo(vm)
else:
enmo = None
if "PITCH" and "ROLL" in stats:
pitch, roll = channel_inference.infer_pitch_roll(x, y, z)
else:
pitch = roll = None
# Infer nonwear and mask those data points in the signal
nonwear_bouts = channel_inference.infer_nonwear_triaxial(
x, y, z, noise_cutoff_mg=noise_cutoff_mg)
for bout in nonwear_bouts:
# Show non-wear bouts in purple
bout.draw_properties = {'lw': 0, 'alpha': 0.75, 'facecolor': '#764af9'}
for channel, channel_name in zip(
[enmo, vm_hpf, pitch, roll, temperature, battery],
["ENMO", "HPFVM", "PITCH", "ROLL", "Temperature", "Battery"]):
if channel_name in stats:
# Collapse the sample data to a processing epoch (in seconds) so data is summarised
epoch_level_channel = channel.piecewise_statistics(
timedelta(seconds=processing_epoch), time_period=tp)[0]
epoch_level_channel.name = channel_name
if channel_name in ["Temperature", "Battery"]:
pass
else:
epoch_level_channel.delete_windows(nonwear_bouts)
epoch_level_channel.delete_windows(missing_bouts)
ts.add_channel(epoch_level_channel)
# collapse binary integrity channel
epoch_level_channel = integrity.piecewise_statistics(
timedelta(seconds=int(processing_epoch)),
statistics=[("binary", ["flag"])],
time_period=tp)[0]
epoch_level_channel.name = "Integrity"
epoch_level_channel.fill_windows(missing_bouts, fill_value=1)
ts.add_channel(epoch_level_channel)
# create and open results files
results_files = [
os.path.join(results_folder, "{}_{}.csv".format(name, filename_short))
for name in names
]
files = [open(file, "w") for file in results_files]
# Write the column headers to the created files
for f in files:
f.write(pampro_utilities.design_file_header(stats) + "\n")
# writing out and plotting results
for epoch, name, f in zip(epochs, names, files):
results_ts = ts.piecewise_statistics(epoch,
statistics=stats,
time_period=tp,
name=id_num)
results_ts.write_channels_to_file(file_target=f)
f.flush()
if name in plots_list:
# for each statistic in the plotting dictionary, produce a plot in the results folder
for stat, plot in plotting_dict.items():
try:
results_ts[stat].add_annotations(nonwear_bouts)
results_ts.draw([[stat]],
file_target=os.path.join(
results_folder,
plot.format(filename_short, name)))
except KeyError:
pass
header["processing_script"] = version
header["analysis_resolutions"] = names
header["noise_cutoff_mg"] = noise_cutoff_mg
header["processing_epoch"] = processing_epoch
header["QC_first_battery_pct"] = first_battery_pct
header["QC_last_battery_pct"] = last_battery_pct
metadata = {**header, **anomalies_dict, **cal_diagnostics}
# write metadata to file
pampro_utilities.dict_write(meta, id_num, metadata)
for c in ts:
del c.data
del c.timestamps
del c.indices
del c.cached_indices
def calibrate(x,y,z, allow_overwrite=True, budget=1000, noise_cutoff_mg=13):
""" Use still bouts in the given triaxial data to calibrate it and return the calibrated channels """
calibration_diagnostics = OrderedDict()
vm = channel_inference.infer_vector_magnitude(x,y,z)
# Get a list of bouts where standard deviation in each axis is below given threshold ("still")
still_bouts = channel_inference.infer_still_bouts_triaxial(x,y,z, noise_cutoff_mg=noise_cutoff_mg, minimum_length=timedelta(minutes=1))
num_still_bouts = len(still_bouts)
num_still_seconds = Bout.total_time(still_bouts).total_seconds()
# Summarise VM in 10s intervals
vm_windows = vm.piecewise_statistics(timedelta(seconds=10), [("generic", ["mean"])], time_period=vm.timeframe)[0]
# Get a list where VM was between 0.5 and 1.5g ("reasonable")
reasonable_bouts = vm_windows.bouts(0.5, 1.5)
num_reasonable_bouts = len(reasonable_bouts)
num_reasonable_seconds = Bout.total_time(reasonable_bouts).total_seconds()
# We only want still bouts where the VM level was within 0.5g of 1g
# Therefore insersect "still" time with "reasonable" time
still_bouts = Bout.bout_list_intersection(reasonable_bouts, still_bouts)
# And we only want bouts where it was still and reasonable for 10s or longer
still_bouts = Bout.limit_to_lengths(still_bouts, min_length = timedelta(seconds=10))
num_final_bouts = len(still_bouts)
num_final_seconds = Bout.total_time(still_bouts).total_seconds()
# Get the average X,Y,Z for each still bout (inside which, by definition, XYZ should not change)
still_x, num_samples = x.build_statistics_channels(still_bouts, [("generic", ["mean", "n"])])
still_y = y.build_statistics_channels(still_bouts, [("generic", ["mean"])])[0]
still_z = z.build_statistics_channels(still_bouts, [("generic", ["mean"])])[0]
# Get the octant positions of the points to calibrate on
occupancy = octant_occupancy(still_x.data, still_y.data, still_z.data)
# Are they fairly distributed?
comparisons = {"x<0":[0,1,2,3], "x>0":[4,5,6,7], "y<0":[0,1,4,5], "y>0":[2,3,6,7], "z<0":[0,2,4,6], "z>0":[1,3,5,7]}
for axis in ["x", "y", "z"]:
mt = sum(occupancy[comparisons[axis + ">0"]])
lt = sum(occupancy[comparisons[axis + "<0"]])
calibration_diagnostics[axis + "_inequality"] = abs(mt-lt)/sum(occupancy)
# Calculate the initial error without doing any calibration
start_error = evaluate_solution(still_x, still_y, still_z, num_samples, [0,1,0,1,0,1])
# Do offset and scale calibration by default
offset_only_calibration = False
calibration_diagnostics["calibration_method"] = "offset and scale"
# If we have less than 500 points to calibrate with, or if more than 2 octants are empty
if len(still_x.data) < 500 or sum(occupancy == 0) > 2:
offset_only_calibration = True
calibration_diagnostics["calibration_method"] = "offset only"
# Search for the correct way to calibrate the data
calibration_parameters = find_calibration_parameters(still_x.data, still_y.data, still_z.data, offset_only=offset_only_calibration)
for param,value in zip("x_offset,x_scale,y_offset,y_scale,z_offset,z_scale".split(","), calibration_parameters):
calibration_diagnostics[param] = value
for i,occ in enumerate(occupancy):
calibration_diagnostics["octant_"+str(i)] = occ
# Calculate the final error after calibration
end_error = evaluate_solution(still_x, still_y, still_z, num_samples, calibration_parameters)
calibration_diagnostics["start_error"] = start_error
calibration_diagnostics["end_error"] = end_error
calibration_diagnostics["num_final_bouts"] = num_final_bouts
calibration_diagnostics["num_final_seconds"] = num_final_seconds
calibration_diagnostics["num_still_bouts"] = num_still_bouts
calibration_diagnostics["num_still_seconds"] = num_still_seconds
calibration_diagnostics["num_reasonable_bouts"] = num_reasonable_bouts
calibration_diagnostics["num_reasonable_seconds"] = num_reasonable_seconds
if allow_overwrite:
# If we do not need to preserve the original x,y,z values, we can just calibrate that data
# Apply the best calibration factors to the data
do_calibration(x, y, z, calibration_parameters)
return (x, y, z, calibration_diagnostics)
else:
# Else we create an independent copy of the raw data and calibrate that instead
cal_x = copy.deepcopy(x)
cal_y = copy.deepcopy(y)
cal_z = copy.deepcopy(z)
# Apply the best calibration factors to the data
do_calibration(cal_x, cal_y, cal_z, calibration_parameters)
return (cal_x, cal_y, cal_z, calibration_diagnostics)
