Table 4 A summary of the three application areas of digital health in COPD

Developing COPD Screening and Diagnosis

Windmon et al. [65]

36 individuals 9 COPD, 9 CHF and 18 CONTROLS

Samsung Galaxy S5 smart-phone a custom voice recording Android application(VoiceRecorder)

cough

features(Zero Crossing Rate (ZCR), Sound Pressure Level (SPL), Interquartile Range (IQR), Percentiles (PER), Mean Absolute Deviation (MAD), Standard Deviation (STD))

Random Forests

SVM k-Nearest Neighbors

Naive Bayes

Sensitivity = 80%

Specificity = 82%

Accuracy = 80.67%

AUC = 83%;

Jhunjhunwala et al. [69]

desktop flow sensor portable thermal mass flowmeter

flowrate, air pressure values

2 features

flowrate, air pressure values

SVM

CNN

LSTM

Zhang et al. [33]

20 individuals (healthy volunteers = 5, asthma = 6, bronchitis = 4, COPD = 5)

smart face mask, a biodegradable self-powered breath sensor

respiratory signals

Various breathing conditions of normal breath, deep breath, fast breath, and cough

26 features

typical bagged decision tree

Accuracy = 93.2%

Abineza et al. [71]

2501 individuals

a pulse oximeter device

baseline, vital, and symptom data and a classification to a COPD exacerbation classes: yes, or not.

pulse rate, Oxygen saturation or SPO2,dyspnea, age, the degree of change of wheezing, the degree of change of shortness of breath and cough, a statement of whether respiratory symptoms wake up the concerned patient at night more than usual.

Logistic Regression,

K-Neighbors,

Random Forest,

XGBoost

and Decision Tree

accuracy = 91.04%

precision = 99.86%

recall = 82.19%

F1 = 90.05%

AUC = 95.8%

D. Bhagya et al. [39]

85 individuals COPD, CHF, and healthy individuals

a speed of sound capnographic sensor HCSR04 ultrasonic sensor laptop

capnographic signal dataset

the capnographic signal

1-D Deformable CNN

accuracy = 92.16%

F score = 0.8542

precision = 0.8597

recall = 0.8488

Identification and Response to COPD Exacerbations

Pereira J et al. [32]

91 individuals

COPD patient

CIDSS(the vital signs prediction module and the early warning score calculation module)

vital signs: SpO2, heart rate, body temperature, SBP and DBP, the number of steps, body fat, energy burned, weight, and height temperature, humidity, wind, and rain PM10, PM2.5

LightGBM

XGBoost

LSTM

BILSTM

GRU

SpO2,LightGBM

RMSE = 0.064668

Heart Rate, GRU

RMSE = 0.110159

Systolic Blood Pressure, GRU

RMSE = 0.130179

Body Temperature, LightGBM

RMSE = 0.058705

Syed Ahmar et al. [39]

100 individuals

The EDGE system(a customized application, a Bluetooth-enabled pulse oximeter, and a secureback-end server)

symptom diary and pulse oximetry data

3 vital signs (heart rate, SpO2, and respiratory rate)

logistic regression

AUC = 0.682,

Specificity = 68 − 36%

sensitivity = 60–80%

A. Dias et al. [52]

58 individuals COPD undergoing LTOT(35 males, 23 females)

GT1M accelerometer (Actigraph LLC)

RT3 accelerometer (Stayhealthy)

Physical activity (PA)

17 features

SD2, SK2, LC2, MC3, AC242, AC243, MF2, MF3, E2, E3, LCF2, LCF3, HLC3, HACC2, HACC3, HLCC2, HLCC3

logarithmic regression (LOG)

support vector machines (SVM)

feed-forward neural network (NN) with 50 neurons in the hidden layer

specificity = 85%,

sensitivity = 100%,

AUC = 90%

Miguel et al. [57]

16 individuals all COPD

an electronic sensor

respiratory sounds

11 features

decision tree forest (DTF)

Accuracy = 87.8%

Sensitivity = 78.1%

Specificity = 95.9%

COPD Patient Monitoring

Juen et al. [28]

38 individuals

28 COPD, CHF and other conditions requiring a pulmonary function test

and 10 subjects without a diagnosed pulmonary disease

smart phones(Motorola Droid Mini, Samsung Galaxy Ace, and LG Optimus Zone)

a middleware software(MoveSense)

Bluetooth Pulse Oximeter(Nonin Onyx II 9560)

medical accelerometers

3329 data points

lap time

9 features

eight spatio-temporal gait parameters, cadence

Shannon entropy, peak frequency

SVM

component extraction approach (CEA)

feature extraction approach (FEA)

Gaussian Process Regression (GPR) model

ANN

controlled 6MWT:

error = 3.23%

free walking:

error = 11.2%

Daniel et al. [46]

18 individuals (COPD and stable chronic respiratory failure)

iPOC(Inogen One G2 POC, an external portable electronic system, a sensor unit, a control unit)

98 time-domain features

2160 epochs of sedentary, 3115 epochs of light activity and 550 epochs of moderate activity

classifiers φ1:9 features

classifiers φ2:14 features

decision trees

Linear Discriminant Analysis (LDA)

Logistic Regression

SVM

classifiers φ1:

Accuracy = 84.01%

F1 = 79.25%

sensitivity = 75.09%

specificity = 85.59%

AUC = 0.88

classifiers φ2:

Accuracy = 96.45%

F1 = 96.98%

sensitivity = 96.76%

specificity = 95.97%

AUC = 0.98

Mireia et al. [72]

59 individuals

40 COPD and 19 non-COPD control

a voice recorder (Model EVIDA L69) the Praat speech analysis software

The Weka Workbench

a primary health care centre (CAP)

ECAP database

Scenario 1:11 speech features

Random Forest

preFEV1

accuracy = 75.0%,

Sensitivity = 81.1%

specificity = 71.0%

postFEV1

accuracy = 73.9%,

Sensitivity = 77.7%

specificity = 71.1%