Application of data-driven technologies for asthma self-management

Abstract

INTRODUCTION: Asthma is a variable chronic condition, affecting around 5.4 million people in the UK and more than 300 million people worldwide. Every 10 seconds in the UK, someone has an asthma attack. Some of these attacks are life-threatening with over 1,400 annual deaths estimated in the UK. Since there is no known cure for asthma, self-management is a key part of patient care; this involves detecting deterioration and taking appropriate action to maintain control and prevent the attack.

To determine a patient’s asthma condition, traditional self-management action plans use symptom scores, sometimes supplemented by peak flow measurements. Keeping track of relief inhaler usage can also help measure asthma control. However, patients may regard this level of monitoring as tedious because it involves high levels of active engagement on their part.

Machine learning and mobile health (mHealth) have the potential to enable a new form of asthma self-management where tedious daily monitoring is minimised, through a combination of passive monitoring and attack prediction.

AIMS AND OBJECTIVES: The aim of my thesis was to progress towards an intelligent mobile health system to support asthma self-management, with four objectives:

1.Benchmark asthma attack predictions which use machine learning algorithms and existing mHealth data.

2.Explore the sub-groups of the asthma population observed in the data.

3.Collect a novel dataset, where asthma patients use connected market-available mHealth devices, which is suitable for machine learning training, for improving existing asthma attack predictions using rich multi-dimensional data.

4.Use the novel dataset to investigate a.the feasibility of asthma patients monitoring with devices, b.the value of objective and passive monitoring data compared to questionnaire data in improving asthma attack predictions, and c.if passive monitoring can directly substitute daily active monitoring.

METHODS: 1.I conducted two data analyses on the Asthma Mobile Health Study (AMHS) (Chan et al., 2017) dataset. The first was a longitudinal study. I benchmarked four supervised learning prediction algorithms (decision tree, logistic regression, naïve Bayes, and support vector machine) to predict asthma attacks based on daily questionnaires. The data were processed with a novel application of the bin-packing algorithm. Features capturing changes over time were extracted during the pre-processing stage and all features were ranked using least absolute shrinkage and selection operator (LASSO).

2.Also using the AMHS data, I conducted a cross-sectional study. I used unsupervised learning (k-means clustering) to identify patient clusters based on markers of asthma attack. I then applied supervised learning (LASSO) to identify the key risk factors associated with each patient cluster, and sensitivity analysis and k-fold cross-validation for internal validation.

3. I conducted the “Mobile device monitoring to inform prediction of asthma attacks” (AAMOS-00) observational study with asthma patients across the UK to collect a novel dataset which included asthma questionnaires and objective monitoring data. Patients who had an asthma attack in the past year were recruited via social media, the Norfolk and Norwich University Hospital, and across the Asthma + Lung UK network. Participants who enrolled into the study begun with one month of daily and weekly questionnaire monitoring, referred to as phase 1 of the study. A maximum of 30 participants who regularly answered the questionnaires in a month were selected to join phase 2 of the study. In phase 2, participants were provided with three smart monitoring devices (smart peak flow meter, smart inhaler, and smartwatch) and continued questionnaire monitoring to collect data for a further six months. Participants also used their smartphone’s location to link with local weather and air quality reports. At the end of phase 2, I collected the participants’ perspectives on the acceptability and utility of the system. Most of the data were collected via a single mobile app, a research platform called Mobistudy. I further developed the platform with the Mobistudy team to integrate all tools in one place.

4.Analysing the AAMOS-00 data, a. I investigated the compliance to passive (smartwatch) and active (asthma diary and peak flow) monitoring over six months. I also explored the usability and acceptability of the monitoring system using questionnaire and written feedback.

b. I compared four sets of models trained on the AAMOS-00 data against the models benchmarked on the AMHS data. The feature processing pipeline and endpoints were the same as the benchmarking analysis. The first set of models also used the same features as the benchmarking analysis, which were all collected via active monitoring.

The second set of models used only passive monitoring data (i.e. smartwatch data, environment data, and relief inhaler usage). The third set of models combined active and passive monitoring data. The fourth combined active and passive monitoring data but excluded peak flow measurements.

c. I paired daily four asthma questions (nocturnal symptoms, daytime symptoms, activity limitation, and relief inhaler usage) with associated passive monitoring signals to investigate whether there were sufficiently strong correlations.

RESULTS: 1. The machine learning predictors developed were explainable and had a good fit to the data. The logistic regression and naïve Bayes models were the best in terms of performance, able to identify weeks leading to an asthma attack with AUC > 0.87 and AUPRC > 0.35, giving a five-fold increase in precision over a random classifier. The most useful features were quick-relief puffs, nocturnal symptoms, frequency of data entry, and day symptoms.

2. Five distinct patient clusters were identified in the AMHS data, ranging from patients with no unscheduled care to patients with hospitalisations and asthma attacks. In addition to nocturnal symptoms, I found that the most important risk factors associated with asthma attacks included previous asthma-related work absenteeism, which was rarely captured systematically in clinical practice. The results were validated at different levels of missing data and had achieved an AUC ≤ 0.85 during k-fold cross-validation.

3. Over 14 months of data collection from April 2021 to June 2022, 32 participants were recruited in phase 1 of the AAMOS-00 observational study and provided 583 patient-days of data with an average of 21 days retention (75% of 28 days). Then 22 participants, selected from phase 1, collected 2,054 patient-days of data in phase 2 with an average of 123 days of retention (67% of 184 days).

4.Using the AAMOS-00 dataset, I found: a.There was no evidence that patients were more compliant to passive smartwatch monitoring than daily questionnaire monitoring. However, the participants faced technical issues with the three monitoring devices. Nonetheless, as evidenced by standard usability questionnaires, the quality of data collection system of the AAMOS-00 study was similar to that of other studies and published apps.

b.The benchmarking of models using the AAMOS-00 data were generally comparable to the benchmark set by models using the AMHS data. The logistic regression model achieved an AUC > 0.79 and AUPRC > 0.39, giving a three-fold increase in precision over a random classifier. Using the passive monitoring data alone gave reasonable performance, AUC > 0.69 and AUPRC > 0.40. When active monitoring data was combined with passive monitoring data, the naïve Bayes model achieved the highest performance, AUC > 0.79 and AUPRC > 0.45. In particular, the most useful features were nocturnal symptoms, relief inhaler usage, and peak flow. c.There was low to negligible correlation between daily asthma questionnaires and the associated passive monitoring signals.

CONCLUSIONS: Machine learning-based asthma attack prediction models using data from mobile health devices have shown strong performance, significantly outperforming random classifiers. Their ability to provide information at an appropriate time to assist patients to make informed decisions is promising. Furthermore, as widely requested in the public, a single mobile app that combines multiple smart monitoring devices and sources of information to support asthma self-management is technically feasible. Analysing data from passive and active monitoring sources revealed that passive data alone could predict some asthma attacks, and passive monitoring devices can enhance current practices of asthma monitoring.

IMPLICATIONS: I found that passive monitoring using mHealth combined with machine learning-based asthma attack prediction is a promising approach to support asthma management. However, so far, passive monitoring is not suitable as a complete substitute for active questionnaire monitoring. In addition, asthma attack prediction models need to be more robust before asthma attack predictors are tested prospectively, as a medical intervention.