Improved Precision of COPD Exacerbation Detection in Night-Time Cough Monitoring

Abstract

Background/Objectives: Targeting individuals with certain characteristics provides improved precision in many healthcare applications. An alert mechanism for COPD exacerbations has recently been validated. It has been argued that its efficacy improves considerably with stratification. This paper provides an in-depth analysis of the cough data of the stratified cohort to identify options for and the feasibility of improved precision in the alert mechanism for the intended patient group. Methods: The alert system was extended using a system complementary to the existing one to accommodate observed rapid changes in cough trends. The designed system was tested in a post hoc analysis of the data. The trend data were inspected to consider their meaningfulness for patients and caregivers. Results: While stratification was effective in reducing misses, the augmented alert system improved the sensitivity and number of early alerts for the acute exacerbation of COPD (AE-COPD). The combination of stratification and the augmented mechanism led to sensitivity of 86%, with a false alert rate in the order of 1.5 per year in the target group. The alert system is rule-based, operating on interpretable signals that may provide patients or their caregivers with better insights into the respiratory condition. Conclusions: The augmented alert system operating based on cough trends has the promise of increased precision in detecting AE-COPD in the target group. Since the design and testing of the augmented system were based on the same data, the system needs to be validated. Signals within the alert system are potentially useful for improved self-management in the target group.

Keywords: COPD; alert; cough; exacerbation; rule-based classification; stratification.

Figure 1

(A) Processing of the night-time cough count C. The counts are mapped onto the B-scale and processed by the cough-based alert mechanism (CAM) to produce alerts A. Details of the CAM are provided in Figure 2. (B) Graph showing the compressive nature of the map $C\to B$, with # denoting the number of coughs observed in a night-time session.

Figure 2

Proposed alert system with the time series B as input, a mapped version of the session cough count C. It processes B in two blocks: the validated unit CAM-V and the proposed fast-response unit CAM-F. The outputs of both systems are merged into alert A by a logical -OR operation (∨). The CAM-V units are as follows: BL—baseline creation; TS—temporal smoother; DU—decision unit. The CAM-F units are as follows: CI—unit to verify consistent increase; TI—unit to quantify total increase; ∧—logical -AND operation.

Figure 3

Illustration of the four possible associations between acute exacerbation (AE) periods and alert trains (represented by vertical arrows). Each AE is preceded by a 14-day early alert window (EAW). (A) Timeline with a missing alert (no start of an alert train during AE or EAW) and a false alert (alert train starting outside of AE and EAW). (B) Timeline with two detected exacerbations—the first one a late alert and the second one an early alert. The horizontal double arrow indicates the lead time.

Figure 4

Examples of raw cough counts (asterisks), smoothed cough counts (black line), and threshold levels (green line) as a function of the monitored day. Alert days are indicated by a coloured dashed line for the validated mechanism (red) and the new one (magenta). The magenta circles provide the data giving rise to the alert in CAM-F: four consecutive increases with a total increase exceeding the threshold of 1 B. The red horizontal line indicates the exacerbation period.

Figure 5

Examples of raw cough counts (asterisks), smoothed cough counts (black dashed lines), and dominant bands (between solid black lines) as a function of the monitored day for four patients from the target group. These graphs illustrate the large variability in the observed patterns across patients, moving from gradual and smooth transitions to sudden jumps.

Figure 6

Examples of raw cough counts (asterisks), smoothed cough counts (black dashed lines), and dominant bands (between solid black lines) as a function of the monitored day for two COPD patients that are not in the target group.