PREDICTIVE MONITORING AND DETECTION OF EXACERBATIONS IN PATIENTS WITH CHRONIC...

Post on 01-Aug-2020

0 views 0 download

transcript

PREDICTIVE MONITORING AND DETECTION OF EXACERBATIONS IN PATIENTS WITH CHRONIC OBSTRUCTIVE PULMONARY DISEASE (COPD)

Peter Gyring

Supervised by

Prof. David CliftonDr. Carmelo Velardo

CONTENT

What is COPD?

What datasets are available?

Brief recap of results from the pilot study

what can be improved, what is lacking?

Goals for the current project (RCT dataset)

Current results

Next steps

WHAT IS COPD?

life-threatening lung disease

long-term decreased airflow

shortness of breath, cough with sputum production

not curable

older terms: emphysema and chronic bronchitis

EXACERBATIONS

Sudden worsening of symptoms

Typically lasts several days

Correlated with worsening of underlying COPD

Early detection and treatment is important

COPD VS ASTHMA

Similar symptoms

Different age of presentation

Different triggers

WHAT DATASETS ARE AVAILABLE?

Two datasets using EDGE-protocol: 1) pilot study, 2) RCT study

Pilot study: 18 patients

RCT study: 110 patients

FEATURES

heart rate

peripheral capillary oxygen saturation, SpO2

respiratory rate

self-reported symptom score

medication score

SYMPTOM SCOREPeter Gyring Chapter 4

Question Range of valuesHow are you feeling today? [0, 5]How is your breathlessness today? [0, 5]How is your wheeze or chest tight-ness today?

[0, 5]

Do you have a cough? yes / noHow is your cough today? [0, 3]Are you coughing up sputum? [0, 4]What colour is your sputum? [White, Brownish]Do you have a cold (such asa runny/blocked nose) or sorethroat?

yes / no

Did you wake up last night due tobreathing problems?

[0, 5]

Table 4.1: Symptom diary questions and their potential responses

Medication score Medication intake0 No medications1 Reliever2 Steroids3 Antibiotics4 Reliever and steroids5 Reliever and antibiotics6 Reliever, steroids and antibiotics

Table 4.2: Medication score values and their corresponding medication intake. Note that thecombination steroids/antibiotics never appears.

4.2.3 Features in the datasets

Both datasets contain the following data, collected at discrete times, and often once daily:pulse rate, peripheral arterial blood oxygen saturation (SpO°2), a self-reported symptomdiary, and a medication score. Additionally, the pilot study contains estimates of the breathingrate, but this is not yet available for the RCT study.

The symptom diary consists of multiple questions which are collectively mapped onto asingle symptom score. The individual questions are summarised in table 4.1.

The medication score, which takes values in the range [0,6], indicates which, if any, med-ications the patient is currently taking. The medications monitored were use of relieverinhaler, steroids, and antibiotics. Table 4.2 shows what medscore values correspond to whatmedications. As can be seen, a value of 4, 5 or 6 indicates that a combination of at least twomedications are in use.

Missing data

Even when data collection takes place, as indicated in the preceding section, some data maybe missing, e.g. due to a faulty pulse oximeter. Table 4.3 summarises the percentage of missingfeature components across all patients.

TODO: in the next subsubsections, introduce each feature in more detail, includingunivariate KDE

23

MEDICATION SCORE

Peter Gyring Chapter 4

Question Range of valuesHow are you feeling today? [0, 5]How is your breathlessness today? [0, 5]How is your wheeze or chest tight-ness today?

[0, 5]

Do you have a cough? yes / noHow is your cough today? [0, 3]Are you coughing up sputum? [0, 4]What colour is your sputum? [White, Brownish]Do you have a cold (such asa runny/blocked nose) or sorethroat?

yes / no

Did you wake up last night due tobreathing problems?

[0, 5]

Table 4.1: Symptom diary questions and their potential responses

Medication score Medication intake0 No medications1 Reliever2 Steroids3 Antibiotics4 Reliever and steroids5 Reliever and antibiotics6 Reliever, steroids and antibiotics

Table 4.2: Medication score values and their corresponding medication intake. Note that thecombination steroids/antibiotics never appears.

4.2.3 Features in the datasets

Both datasets contain the following data, collected at discrete times, and often once daily:pulse rate, peripheral arterial blood oxygen saturation (SpO°2), a self-reported symptomdiary, and a medication score. Additionally, the pilot study contains estimates of the breathingrate, but this is not yet available for the RCT study.

The symptom diary consists of multiple questions which are collectively mapped onto asingle symptom score. The individual questions are summarised in table 4.1.

The medication score, which takes values in the range [0,6], indicates which, if any, med-ications the patient is currently taking. The medications monitored were use of relieverinhaler, steroids, and antibiotics. Table 4.2 shows what medscore values correspond to whatmedications. As can be seen, a value of 4, 5 or 6 indicates that a combination of at least twomedications are in use.

Missing data

Even when data collection takes place, as indicated in the preceding section, some data maybe missing, e.g. due to a faulty pulse oximeter. Table 4.3 summarises the percentage of missingfeature components across all patients.

TODO: in the next subsubsections, introduce each feature in more detail, includingunivariate KDE

23

MEDICATION SCORE

Peter Gyring Chapter 4

Question Range of valuesHow are you feeling today? [0, 5]How is your breathlessness today? [0, 5]How is your wheeze or chest tight-ness today?

[0, 5]

Do you have a cough? yes / noHow is your cough today? [0, 3]Are you coughing up sputum? [0, 4]What colour is your sputum? [White, Brownish]Do you have a cold (such asa runny/blocked nose) or sorethroat?

yes / no

Did you wake up last night due tobreathing problems?

[0, 5]

Table 4.1: Symptom diary questions and their potential responses

Medication score Medication intake0 No medications1 Reliever2 Steroids3 Antibiotics4 Reliever and steroids5 Reliever and antibiotics6 Reliever, steroids and antibiotics

Table 4.2: Medication score values and their corresponding medication intake. Note that thecombination steroids/antibiotics never appears.

4.2.3 Features in the datasets

Both datasets contain the following data, collected at discrete times, and often once daily:pulse rate, peripheral arterial blood oxygen saturation (SpO°2), a self-reported symptomdiary, and a medication score. Additionally, the pilot study contains estimates of the breathingrate, but this is not yet available for the RCT study.

The symptom diary consists of multiple questions which are collectively mapped onto asingle symptom score. The individual questions are summarised in table 4.1.

The medication score, which takes values in the range [0,6], indicates which, if any, med-ications the patient is currently taking. The medications monitored were use of relieverinhaler, steroids, and antibiotics. Table 4.2 shows what medscore values correspond to whatmedications. As can be seen, a value of 4, 5 or 6 indicates that a combination of at least twomedications are in use.

Missing data

Even when data collection takes place, as indicated in the preceding section, some data maybe missing, e.g. due to a faulty pulse oximeter. Table 4.3 summarises the percentage of missingfeature components across all patients.

TODO: in the next subsubsections, introduce each feature in more detail, includingunivariate KDE

23

Green

Green/Amber

Red

PERIOD OF DATA COLLECTION

RCT studyPilot study

EXAMPLE SCATTER PLOTS

EXAMPLE SCATTER PLOTS

PILOT STUDY

Novelty detection algorithms

Medication score 0-3 is good, 4-6 is bad

AUC up to 0.91

PILOT STUDY

is also shown in the figure. Based on a specific threshold (95% in this case), the data points with univariate alerts are also indicated. It can be seen that each of the variable generates alerts during the medication event, but the performance of the Parzen windows based method (multivariate novelty score) is superior to the rest as no false alerts are generated.

Table 3 Comparison of area under curve (AUC) for each of the personalized alerting methods used in the study

Method AUC

Multivariate (with respiratory rate)

0.91

Multivariate (without respiratory rate)

0.88

Respiratory rate 0.88

Pulse 0.84

SpO2 0.81

Symptoms score 0.76

Figure 3 shows the ROC curves for each of the univariate (SpO2, pulse, symptoms score and respiratory rate) and the multivariate methods (with and without the addition of respiratory rate). To compare the overall performance of each method, we use the AUCs (Table 3). Multivariate novelty detection using Parzen windows improves performance (AUC rises to 0.88). Respiratory rate on its own performs as well as the combination of the other three variables. The addition of respiratory rate to the multivariate method leads to further improvement, giving an AUC of 0.91.

Figure 3 The ROC curves for the univariate and the multivariate alerting algorithms (with and without the addition of respiratory rate)

Since it is not feasible to obtain an objective measure of lung infection, we relied on the self-reported medication events as exacerbation indicators. However, relying on patients to identify exacerbation periods may be imprecise. Some patients could miss an event, while others might take medication even during normal periods. Future work could use information from healthcare professionals (e.g. hospital admissions) to help identify exacerbations more accurately.

IV. CONCLUSION

We evaluated both univariate and multivariate approaches using data from 16 COPD patients acquired over a period of six months using an m-Health system based on a tablet computer and a pulse oximeter. We have described a novelty detection framework for generating personalized alerts to prioritize patients for clinical review. We found that using a multivariate approach which includes estimates of respiratory rate gives the best result using retrospective analysis on pilot data. Future work involves the use of the multivariate approach in a randomized control trial with 165 COPD patients.

ACKNOWLEDGMENT

The authors would like to acknowledge the COPD clinical team, and the patients who participated in the study.

REFERENCES

[1] WHO, World Health Statistics 2008, World Health Organization, 2008 [2] National Institute for Health and Care Excellence, Chronic obstructive

pulmonary disease, NICE, 2010 [3] EWMA Bischoff, R Akkermans, J Bourbeau, C van Weel, JH

Vercoulen, TRJ Schermer, “Comprehensive self-management and routine monitoring in chronic obstructive pulmonary disease patients in general practice: randomised controlled trial”, BMJ: British Medical Journal, vol. 345, 2012

[4] JA Wedzicha, J Vestbo, “Can patients with COPD self-manage?”, The Lancet, vol. 380, no. 9842, pp. 624–625, 2012

[5] AM Yañez, D Guerrero, R Pérez de Alejo, F Garcia-Rio, JL Alvarez-Sala, M Calle-Rubio, R Malo de Molina et al. “Monitoring breathing rate at home allows early identification of COPD exacerbations”, Chest vol. 142, no. 6, pp. 1524-1529, 2012

[6] L Tarassenko, A Nairac, N Townsend, I Buxton, P Cowley, “Novelty detection for the identification of abnormalities”, International Journal of Systems Science, vol. 31, no. 11, pp. 1427–1439, 2000

[7] L Tarassenko, P Hayton, N Cerneaz, M Brady, “Novelty detection for the identification of masses in mammograms”, in Artificial Neural Networks, Fourth International Conference on. IET, pp. 442–447, 1995

[8] L Tarassenko, A Hann, D Young, “Integrated monitoring and analysis for early warning of patient deterioration”, British Journal of Anaesthesia, vol. 97, no. 1, pp. 64–68, 2006

[9] MAF Pimentel, DA Clifton, L Clifton, L Tarassenko, “A review of novelty detection”, Signal Processing, vol. 99, pp. 215-249, 2014

[10] A Farmer, C Toms, M Hardinge, V Williams, H Rutter, L Tarassenko, “Self-management support using an internet-linked tablet computer (the EDGE platform)-based intervention in chronic obstructive pulmonary disease: protocol for the EDGE-COPD randomised controlled trial”, BMJ open, vol. 4, no. 1, pp. e004437, 2014

[11] WS Johnston, Y. Mendelson. “Extracting breathing rate information from a wearable reflectance pulse oximeter sensor”, 26th Annual International Conference of the IEEE, vol. 2, pp. 5388-5391. IEEE, 2004

[12] DJ Meredith, D Clifton, P Charlton, J Brooks, CW Pugh, L Tarassenko. "Photoplethysmographic derivation of respiratory rate: a review of relevant physiology", Journal of medical engineering & technology vol. 36, no. 1, pp. 1-7, 2012

[13] SG Fleming, and L Tarassenko. "A Comparison of Signal Processing Techniques for the Extraction of Breathing Rate from the Photoplethysmogram", International Journal of Biomedical Sciences vol. 2, no. 4, 2007

[14] E Parzen “On estimation of a probability density function and mode”, Annals of mathematical statistics vol. 33, pp. 1065–1076, 1962

[15] CM Bishop, Pattern recognition and machine learning, vol. 1, Springer New York, 2006

[16] SA Shah, Vital sign monitoring and data fusion for paediatric triage, DPhil dissertation, Dept. Eng. Science, University of Oxford, UK, 2012

3167

WHAT IS MISSING?

Bigger dataset

Best choice of exacerbation ‘truth’

Population-based algorithms

What are the best predictors?

Segmentation of patient-population?

Seasonal effects?

RCT STUDY

Population-based classification, how well can we do?

How should an exacerbation be defined?

What are the best predictors?

POPULATION-BASED CLASSIFICATION

Peter Gyring Chapter 4

Question Range of valuesHow are you feeling today? [0, 5]How is your breathlessness today? [0, 5]How is your wheeze or chest tight-ness today?

[0, 5]

Do you have a cough? yes / noHow is your cough today? [0, 3]Are you coughing up sputum? [0, 4]What colour is your sputum? [White, Brownish]Do you have a cold (such asa runny/blocked nose) or sorethroat?

yes / no

Did you wake up last night due tobreathing problems?

[0, 5]

Table 4.1: Symptom diary questions and their potential responses

Medication score Medication intake0 No medications1 Reliever2 Steroids3 Antibiotics4 Reliever and steroids5 Reliever and antibiotics6 Reliever, steroids and antibiotics

Table 4.2: Medication score values and their corresponding medication intake. Note that thecombination steroids/antibiotics never appears.

4.2.3 Features in the datasets

Both datasets contain the following data, collected at discrete times, and often once daily:pulse rate, peripheral arterial blood oxygen saturation (SpO°2), a self-reported symptomdiary, and a medication score. Additionally, the pilot study contains estimates of the breathingrate, but this is not yet available for the RCT study.

The symptom diary consists of multiple questions which are collectively mapped onto asingle symptom score. The individual questions are summarised in table 4.1.

The medication score, which takes values in the range [0,6], indicates which, if any, med-ications the patient is currently taking. The medications monitored were use of relieverinhaler, steroids, and antibiotics. Table 4.2 shows what medscore values correspond to whatmedications. As can be seen, a value of 4, 5 or 6 indicates that a combination of at least twomedications are in use.

Missing data

Even when data collection takes place, as indicated in the preceding section, some data maybe missing, e.g. due to a faulty pulse oximeter. Table 4.3 summarises the percentage of missingfeature components across all patients.

TODO: in the next subsubsections, introduce each feature in more detail, includingunivariate KDE

23

Green

Ignore

Red

POPULATION-BASED CLASSIFICATION

110 patients

80 patients used for 10-fold cross-validation

Pick the best hyper parameter

remaining 30 patients used for testing of resulting algorithm

repeat 50 times

LOGISTIC REGRESSION EXAMPLE

COMPARISON OF ALGORITHMS

green: medscore 0

red: medscore 4-6

Algorithm Median AUC 1st quartile AUC 3rd quartile AUC

SVM (rbf kernel) 0.64 0.63 0.67

RDA 0.64 0.62 0.66

K-Nearest 0.65 0.62 0.68

Logistic Regression 0.65 0.62 0.67

Table 2.1: Summary of baseline AUC results, including median and 1st/3rd quartiles.

(a) All ROC curves. (b) Median and quartiles ROC.

(c) Hyperparameters (d) Distribution of AUCs

Figure 2.1: SVM performance details across 50 experiments. In plot (c), red x’s mark thehyperparameter grid search, whereas blue circles mark chosen hyperparamers.Circle diameter is proportional to how many times (out of 50) the given hyperpa-rameter was chosen.

3

WHAT CONSTITUTES AN EXACERBATION?

Keep green fixed as medication score 0

Red = 1-6: AUC 0.60

Red = 4-6: AUC 0.64

Red = 3-6: AUC 0.64

Red = 3,5,6: AUC 0.67

Red = 6: AUC 0.71

Peter Gyring Chapter 4

Question Range of valuesHow are you feeling today? [0, 5]How is your breathlessness today? [0, 5]How is your wheeze or chest tight-ness today?

[0, 5]

Do you have a cough? yes / noHow is your cough today? [0, 3]Are you coughing up sputum? [0, 4]What colour is your sputum? [White, Brownish]Do you have a cold (such asa runny/blocked nose) or sorethroat?

yes / no

Did you wake up last night due tobreathing problems?

[0, 5]

Table 4.1: Symptom diary questions and their potential responses

Medication score Medication intake0 No medications1 Reliever2 Steroids3 Antibiotics4 Reliever and steroids5 Reliever and antibiotics6 Reliever, steroids and antibiotics

Table 4.2: Medication score values and their corresponding medication intake. Note that thecombination steroids/antibiotics never appears.

4.2.3 Features in the datasets

Both datasets contain the following data, collected at discrete times, and often once daily:pulse rate, peripheral arterial blood oxygen saturation (SpO°2), a self-reported symptomdiary, and a medication score. Additionally, the pilot study contains estimates of the breathingrate, but this is not yet available for the RCT study.

The symptom diary consists of multiple questions which are collectively mapped onto asingle symptom score. The individual questions are summarised in table 4.1.

The medication score, which takes values in the range [0,6], indicates which, if any, med-ications the patient is currently taking. The medications monitored were use of relieverinhaler, steroids, and antibiotics. Table 4.2 shows what medscore values correspond to whatmedications. As can be seen, a value of 4, 5 or 6 indicates that a combination of at least twomedications are in use.

Missing data

Even when data collection takes place, as indicated in the preceding section, some data maybe missing, e.g. due to a faulty pulse oximeter. Table 4.3 summarises the percentage of missingfeature components across all patients.

TODO: in the next subsubsections, introduce each feature in more detail, includingunivariate KDE

23

BEST PREDICTORS?

Full symptom score

Sleep last night?

Sputum production?

Heart rate

How is your breathlessness?

NEXT STEPS

Add breathing rate

Novelty detection algorithms

pointwise

timeseries