Application of artificial intelligence in quantifying lung deposition dose of black carbon in people with exposure to ambient combustion particles

Abstract

Background: Understanding lung deposition dose of black carbon is critical to fully reconcile epidemiological evidence of combustion particles induced health effects and inform the development of air quality metrics concerning black carbon. Macrophage carbon load (MaCL) is a novel cytology method that quantifies lung deposition dose of black carbon, however it has limited feasibility in large-scale epidemiological study due to the labor-intensive manual counting.

Objective: To assess the association between MaCL and episodic elevation of combustion particles; to develop artificial intelligence based counting algorithm for MaCL assay.

Methods: Sputum slides were collected during episodic elevation of ambient PM_2.5 (n = 49, daily PM_2.5 > 10 µg/m³ for over 2 weeks due to wildfire smoke intrusion in summer and local wood burning in winter) and low PM_2.5 period (n = 39, 30-day average PM_2.5 < 4 µg/m³) from the Lovelace Smokers cohort.

Results: Over 98% individual carbon particles in macrophages had diameter <1 µm. MaCL levels scored manually were highly responsive to episodic elevation of ambient PM_2.5 and also correlated with lung injury biomarker, plasma CC16. The association with CC16 became more robust when the assessment focused on macrophages with higher carbon load. A Machine-Learning algorithm for Engulfed cArbon Particles (MacLEAP) was developed based on the Mask Region-based Convolutional Neural Network. MacLEAP algorithm yielded excellent correlations with manual counting for number and area of the particles. The algorithm produced associations with ambient PM_2.5 and plasma CC16 that were nearly identical in magnitude to those obtained through manual counting.

Impact statement: Understanding lung black carbon deposition is crucial for comprehending health effects of combustion particles. We developed "Machine-Learning algorithm for Engulfed cArbon Particles (MacLEAP)", the first artificial intelligence algorithm for quantifying airway macrophage black carbon. Our study bolstered the algorithm with more training images and its first use in air pollution epidemiology. We revealed macrophage carbon load as a sensitive biomarker for heightened ambient combustion particles due to wildfires and residential wood burning.

Keywords: Artificial intelligence; Combustion-emitted particulate matter; Lung deposition dose; Macrophage carbon load.

Figure 1.. Pipeline for the Development of the Machine-Learning algorithm for Engulfed cArbon Particles (MacLEAP) algorithm.

A cascaded approach was adopted to develop macrophage and carbon particle models using the Mask R-CNN. A subset (n=1043) of sputum images from 66 LSC subejcts were used to improve the MacLEAP algorithm. A total of 2297 macrophages were masked manually on these 1043 projection images to enhance the macrophage model. A total of 2364 black carbon particles were masked manually on 450 cropped macrophages to develop the black carbon model. The optimized algorithm was then applied to the rest 413 sputum images from the 66 subjects and all 333 sputum images from the 22 subjects.

Figure 2.. Episodic elevation of PM_2.5 and areas of the particles in macrophages.

Large differences in PM_2.5 levels were observed up to 2 months prior to sputum collection, however differences reduced for 3 months and beyond (A). Over 98% of engulfed individual particles had diameters less than 1 μm with 78.6% between 200 and 500 nm (C). 7d average of PM_2.5 was less than 5 μg/m³ for low PM_2.5 group, but ranged from 6 to 25 μg/m³ for high PM_2.5 group. Each 5 μg/m³ increase in 7d PM_2.5 average was associated with 48% and 30% increase in AoPs scored manually (B, P=0.0009) or by MacLEAP (D, P=0.0024), respectively. D: day; M: month.

Figure 3.. MaCL and lung injury biomarker CC16.

Each inter-quartile range increase in AoPs scored manually (A, 0.134 μm²) or by MacLEAP (B, 0.195 μm²) was associated with 19% and 22% reduction in plasma CC16 levels, respectively. Higher cut points for defining macrophages with large carbon occupancy resulted in stronger associations between %cells with higher carbon load and plasma CC16 levels (C).

Figure 4.. Correlations between manual and MacLEAP counting for NoPs and AoPs.

A and D summarized the correlations of AoPs and NoPs between manual and MacLEAP counting for 1456 sputum images from the 66 subjects. The optimized algorithm was then applied to the 333 sputum images from 22 subjects as an independent validation set (B and E). C and F demonstrated the correlations of AoPs and NoPs between manual and MacLEAP counting for all 1789 sputum images from all 88 subjects. Performance was evaluated using individual-based NoPs and AoPs. Red lines represented the 1:1 identity line. Blue dotted lines were the linear regression lines.