The diagnostic accuracy of chest Xray screening for silicosis: A systematic review, meta-analysis and modelling study

Silicosis, Pneumoconiosis, Silicon Dioxide, Quartz, Mining, 2 Industry, Occupational Disease, Occupational Exposure, Respiratory, 3 Computed Tomography, Radiograph, X-ray 4 5 Word counts 6

words: 250 words 7 Manuscript: 3182 words 8 9

37 10 11 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 3

1 2

Chest Xray (CXR) is widely used for silicosis diagnosis, despite concerns 3 regarding sensitivity. We investigated the diagnostic accuracy of CXR for silicosis 4 screening compared to computed tomography (CT), high-resolution CT (HRCT) and 5 autopsy, and modelled the relationship between CXR sensitivity and disease 6 severity. 7 8

Medline, Embase, Scopus, and Web of Science databases were searched 9 on 2nd July 2024 (Prospero registration: CRD42024513830). Meta-analyses were 10 performed by reference standard and at increasing reference test severity cut-offs. 11 The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool 12 assessed risk of bias. In scenarios of fixed and relative sensitivity, according to 13 disease severity, we estimated missed silicosis cases and the number needed to 14 screen (NNS) in hypothetical populations of low (5%), medium (15%) and high (30%) 15 silicosis prevalence. 16 17

Twenty studies included 2156 participants and 1105 silicosis cases. CXR 18 had moderate sensitivity (0.76; 95% confidence interval (CI): 0.63-0.86, I2=84%) and 19 high specificity (0.89, 95% CI: 0.77-0.95, I2=57%) compared to HRCT in 12 studies, 20 and low sensitivity (0.50, 95% CI: 0.45-0.55, I2=0%) and high specificity (0.91, 95% 21 CI: 0.87-0.93, I2=20%) compared to autopsy in two studies. CXR sensitivity 22 increased with higher reference test severity cut-offs. Clinically relevant numbers of 23 cases were missed in fixed and relative sensitivity scenarios; increased prevalence 24 and less severe disease resulted in more missed cases and a lower NNS. 25 26

Silicosis severity and reference test type both plausibly influence CXR 27 sensitivity. Assuming either fixed or relative sensitivity results in missed silicosis 28 cases. Judicious HRCT screening is likely to improve case detection. 29 30 31 32 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 4 What is already known on this topic 1 It is widely understood that Chest Xray (CXR) underdiagnoses silicosis compared to 2 more accurate methods, such as high resolution computed tomography (HRCT) and 3 autopsy. 4 5 What this study adds 6 Our systematic review and meta-analysis demonstrated that the sensitivity of CXR 7 was lowest when compared to autopsy (50%), followed by HRCT (76%). This 8 difference may be explained by the increased accuracy of autopsy as a reference 9 test. Another potential explanation for differences between study results could be 10 that – because severe silicosis is more easily diagnosed by CXR – studies with a 11 higher proportion of severe disease recorded higher sensitivity results. Importantly, 12 regardless of whether differences between studies are explained by different 13

test modalities or the proportion of severe disease, when modelled among 14 a population of silica-exposed workers, many silicosis cases are missed. 15 16 How this study might affect research, practice or policy 17 This study suggests the careful implementation of HRCT screening for silicosis 18 would improve case detection. 19 20 21 22 23 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 5

1 2 Silicosis is a preventable yet incurable occupational fibrotic lung disease caused by 3 the inhalation of respirable crystalline silica (RCS) dust. Recent reports of the high 4 morbidity and mortality of artificial stone silicosis are reminders of the potentially 5 devastating effects of this disease[1,2]. The likely underdiagnosis of silicosis and the 6 lack of clarity around modelling methods mean the global burden of disease is 7 unknown and likely higher than currently reported estimates[3,4]. 8 9 Silicosis diagnosis relies on positive Chest Xray (CXR) and/or Computed 10 Tomography (CT) findings in conjunction with a compatible exposure history. The 11 International Labour Organisation (ILO) classification system[5] is one system that 12 provides a standardised approach to screening for possible silicosis cases in at-risk 13 populations. The 12-point, stepwise classification describes the severity and 14 profusion of small-opacities, along with other relevant findings. In UK and US 15 practice, a profusion cut off of > 1/0 often qualifies as a positive screening result[6,7]; 16 CT is normally reserved for cases with an abnormal CXR. Due to availability and 17 cost, for most workers globally, CXR is the only test performed to diagnose silicosis. 18 In practice, therefore, a positive screening CXR in an exposed worker will be 19 classified as a case of silicosis. In many settings, the expertise required to apply the 20 ILO classification precludes its use, relying instead on a simple radiographic 21 diagnosis. 22 23 The concern that some silicosis may not be visible on CXR, sometimes termed 24 “subradiological silicosis”, dates back to autopsy studies of South African miners, 25 which demonstrated a sensitivity of 31% at a CXR >= 1/1 ILO cut-off[8,9]. More 26 recently, among artificial stone workers in Australia, a CXR sensitivity of 48% was 27 reported compared to HRCT[10]. These enduring concerns are changing practice. A 28 2019 Royal Australian and New Zealand College of Radiologists Position Statement 29 strongly recommended CT scanning as the “primary imaging modality for screening 30 exposed workers”. However, the potential limitations of CXR have not been 31 systematically reviewed or quantified[11]. In practice, and for the purposes of this 32 paper, HRCT and reconstructed multidetector CT (MDCT) images are collectively 33 referred to as HRCT. 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 6 1 It is intuitive that CXR sensitivity is related – at least in part – to severity of disease, 2 such that CXRs with a greater severity of disease have a higher sensitivity. In South 3 African miners, CXR sensitivity increased from 38% in the subgroup with “slight” 4 disease at autopsy, to 83% in those with “marked” disease[9]. The counterfactual 5 assumes that sensitivity does not vary with silicosis severity. These two relationships 6 are illustrated by directed acyclic diagrams in Figure 1A and 1B. 7 8 Quantifying the sensitivity of CXR compared to high-resolution computed 9 tomography (HRCT) informs the debate on the role of HRCT in silicosis 10 screening[12]. Understanding whether silicosis severity may modify CXR sensitivity 11 is helpful clinically and programmatically. Understanding the impact of the 12 counterfactual scenarios of fixed and relative sensitivity, according to silicosis 13 severity, on the number missed cases may be helpful in identifying priority 14 populations for HRCT screening. 15 16 We therefore performed a systematic review and meta-analysis of the sensitivity and 17 specificity of CXR compared to computed tomography (CT), HRCT and autopsy 18

standards. We then modelled how sensitivity and specificity of CXR 19 changes according to severity of underlying disease. We used this model to estimate 20 the number of cases missed and number needed to screen (NNS) in example 21 populations of silica-exposed individuals. 22 23 24 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 7

1 2 One reviewer (AD) searched Medline, Embase, Scopus and Web of Science 3 databases from inception up to 2nd July 2024 (search strategy in online 4 supplementary materials 1) and reviewed bibliographies of included studies 5 (Prospero ID: CRD42024513830). 6 7 Study inclusion criteria were: (1) cohort, cross-sectional, or case-control study 8 design; (2) Population >18 years-old with occupational RCS exposure; (3) CXR used 9 as the index test for assessing silicosis, using the ILO classification or equivalent; (4) 10 CT, HRCT, or autopsy used as the reference standard for assessing silicosis; (5) 11 sufficient information to construct 2x2 tables. In the case of a high risk of duplicated 12 data, the most comprehensive study was retained. 13 14 Title and abstract screening and full-text review was performed in parallel by two 15 reviewers (AD, PH) using Covidence. Data was extracted by a single reviewer (AD) 16 and then checked (PH). Disagreements were resolved through consensus. 17 Methodological quality was assessed using the Quality Assessment of Diagnostic 18 Accuracy Studies-2 (QUADAS-2) tool[13] and entered into RevMan. Extraction tools 19 are presented in online supplementary materials 2 and description of our application 20 of the QUADAS-2 tool in supplementary materials 3. 21 22 Meta-analysis 23 24 Our pre-specified CXR meta-analysis of sensitivity and specificity of CXR for silicosis 25 included studies using the ILO CXR cut-off of ≥ 1/0 or equivalent for the index 26 test[14]. The HRCT group included reconstructed MDCT scans, all of which were 27 labelled as HRCT in their respective studies. Post-hoc sensitivity analyses included 28 studies with other CXR cut-offs and restriction to studies without a high/unclear risk 29 of patient selection or index test bias. 30 31 To demonstrate whether CXR diagnostic accuracy varied according to silicosis 32 severity, we repeated our meta-analyses at increasing reference test severity cut-33 offs. We separated studies by radiology (HRCT and CT) and autopsy reference 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 8 standards. Although recent consensus classifications are available[15], no single 1 HRCT/CT or autopsy classification was consistently applied. We therefore 2 considered four of seven studies that used a 4-level scale (similar to the ILO CXR 3 system) as equivalent and mapped the remainder to that scale (details in online 4 supplement materials 4). 5 6 A bivariate generalized linear mixed model (GLMM) with logit transformation was 7 used to calculate pooled sensitivity and specificity estimates using the “meta” 8 package in R (Version 4.3.2). The bivariate GLMM approach accounts for the 9 correlation between sensitivity and specificity and is preferred when 2x2 cell counts 10 are low[16]. Heterogeneity was assessed using the I2 statistic. 11 12 Measures of impact 13 14 We performed a mixed-effects linear meta-regression with the exposure of the ratio 15 of ILO category ≥ 2/1 relative to ≥ 1/0 silicosis and the outcome of CXR ≥ 1/0 16 sensitivity compared to HRCT. ILO category ≥ 2/1 was used as it represents major 17 category 2 disease and provides a robust numerator. Using this model, and to 18 demonstrate the impact of both fixed and relative sensitivity scenarios, we estimated 19 the sensitivity of CXR for silicosis among hypothetical populations of 1000 silica-20 exposed workers with low (5%), medium (15%) and high (30%) prevalences of CXR 21 ILO ≥ 1/0 category silicosis. We also assumed two fixed proportions of less severe 22 disease (20% CXR ILO ≥ 2/1 relative to ILO >1/0 silicosis) and more severe disease 23 (40% CXR ILO ≥ 2/1 relative to ILO >1/0 silicosis). In each population, we then 24 calculated the number of missed cases that would be identified by HRCT and the 25 number needed to screen using HRCT to detect a single extra case of silicosis. A 26 sensitivity analysis used the prevalence of CXR (non-reference confirmed) CXR ILO 27 category ≥ 2/1 silicosis as the meta-regression exposure. 28 29 All code and data are publicly available at 30 https://github.com/pjhowlett/da_silic_cxr/tree/main. 31 32 Human Research Ethics Committee approval was not required as only previously 33 published data was used. 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 9

1 2 From a total of 825 title and abstracts, 47 studies underwent full-text screening and 3 20 were included (Figure 2). Fifteen studies used HRCT, three used CT and two 4 used autopsy as a reference standard. A total of 2156 participants (HRCT 1207, CT 5 151, Autopsy 798) were included, of whom 1105 (HRCT 587, CT 112, Autopsy 406) 6 were identified by the reference test to have silicosis. There was at least moderate 7 agreement for title and abstract and full-text reviews; Cohen’s Kappa 0.65 and 0.45, 8 respectively. 9 10 All studies represented mining and non-mining industries in middle or high-income 11 economies. The median prevalence of reference test silicosis was 63% (range 16-12 99%) [17–24,15,25–32,8,33,34] (online supplementary table 1). Study size ranged 13 from 11 to 557 participants. Exposures ranged from 7.6 to 31 years; dental 14 technician, sandblaster and artificial stone worker studies had median exposures of 15 <15 years[25,28–30]. Autopsy and CT studies were either performed or published 16 before HRCT in most cases, although HRCT publication dates ranged from 1991 to 17 2024. 18 19 Among 15 studies with HRCT as the reference test, CXR sensitivity ranged from 20 0.18 (95% CI 0.02, 0.52) to 0.95 (95% CI 0.76, 1.00) while specificity ranged from 21 0.29 (95% CI 0.04, 0.71) to 1.00 (95% CI 0.99, 1.00) (Figure 3A). For two studies 22 with autopsy as the reference test, CXR sensitivity was 0.50 (95% CI 0.39, 0.61) and 23 0.50 (95%CI 0.44, 0.56), while specificity was 0.89 (95% CI 0.84, 0.93) and 0.93 24 (95% CI 0.87, 0.96) (Figure 3B). For three studies with CT as the reference test, 25 CXR sensitivity ranged from 0.94 (95% CI 0.73, 1.00) to 1.00 (95% CI 0.92, 1.00) 26 while specificity ranged from 0.50 (95% CI 0.21, 0.79) to 1.00 (95% CI 0.48, 1.00) 27 (supplementary figure 1). 28 29 Quality Assessment 30 31 Eight studies had unclear bias and three had high risk of bias regarding patient 32 selection (online supplementary figure 2). Unclear or high risk of bias in the index (6 33 studies) or reference (4 studies) test was most commonly due to poor detailing of 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 10 test blinding. Six studies had unclear bias and three had high risk of bias for flow and 1 timing, due to flow concerns (2 studies) or unknown or inappropriate intervals 2 between the index and reference test (7 studies). Five studies had high or unclear 3 applicability, mainly related to different ILO CXR cut-offs. 4 5 Meta-analysis 6 7 Pooled sensitivity for CXR with a HRCT reference standard was 0.76 (95% CI 0.63, 8 0.86; n = 12), with a pooled specificity of 0.89 (95% CI 0.77, 0.95; n = 12) (Figure 9 3A). Heterogeneity was substantial for sensitivity (I2 = 84%) and moderate for 10 specificity (I2 = 57%). Compared to an autopsy reference standard, pooled sensitivity 11 and specificity for CXR were 0.50 (95% CI 0.45, 0.55; n = 2) and 0.91 (95% CI 0.87, 12 0.93; n = 2), with low heterogeneity (I2 = 0% and I2 = 20%, respectively) (Figure 3B). 13 Compared to a CT reference standard, pooled sensitivity and specificity for CXR 14 were 0.97 (95% CI 0.86, 0.99; n = 2) and 0.69 (95% CI 0.20, 0.95; n = 2), with low (I2 15 = 0%) and moderate (I2 = 57%) heterogeneity, respectively (supplementary figure 1). 16 The sensitivity analyses, which included different ILO cut-offs or the removal of 17 studies with a high/unclear risk of patient selection bias or index test bias, did not 18 significantly alter our results (online supplementary figures 3 and 4). 19 20 CXR demonstrated increased sensitivity at higher grade cut-offs of silicosis in the 21

test (online supplementary figures 5 and 6). Sensitivity was 0.86 (95% CI 22 0.72, 0.93; I2 = 47%; n = 4) compared to combined HRCT and CT test category 1 23 disease or greater. All participants bar one with reference level 2 or higher disease 24 on CT or HRCT were identified correctly by CXR. For autopsy reference standards, 25 sensitivity increased from 0.72 (95% CI 0.27, 0.95; I2 = 89%; n = 3) at the lowest cut-26 off to 0.83 (95% CI 0.70, 0.91; I2 = 0%; n = 3) at the highest. Correspondingly, 27 specificity showed a decreasing trend across both radiological and autopsy 28

standards with increasing grade of cut off. 29 30 Missed cases and number needed to screen 31 32 Our meta-regression estimated a 1.15% (95% CI 0.44, 1.86) increase in sensitivity 33 for each 1% increase in the ratio of ILO category ≥ 2/1 relative to ≥ 1/0 silicosis. When 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 11 there were no ILO category ≥ 2/1 silicosis cases (i.e. the ratio was zero) the 1 sensitivity of CXR was 16%. The R2 value was 63% and heterogeneity substantial (I2 2 = 86%) (Figure 4). 3 4 In a low prevalence scenario (5% or 50 silicosis cases per 1000 exposed) (Table 1 5 and supplementary figure 7A) and under the assumption of fixed CXR sensitivity, the 6 number of missed cases per 1000 exposed was 16 cases and the NNS was 62 7 persons. Under the assumption of relative sensitivity, the number of cases missed 8 was higher and the NNS lower; when we assumed a lower proportion of severe 9 disease (20% cases > ILO 2/1 relative to >1/0 silicosis), 79 cases per 1000 exposed 10 were missed with NNS of 13 persons. When we assumed severe disease was more 11 common (40% cases > ILO 2/1 relative to >1/0 silicosis) 31 cases per 1000 person 12 were missed with a NNS of 32 persons. Both increasing prevalence of silicosis and a 13 higher proportion of severe disease led to a higher number of missed cases and, 14 hence, a lower NNS. Our sensitivity analysis, using the prevalence of CXR ILO >2/1, 15 resulted in similar estimates of missed cases and NNS (supplementary table 2, 16 supplementary figures 7B and 8). 17 18 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 12 Table 1. Missed cases and number needed to screen in a silica-exposed population. 1 Calculated using the metaregression model with exposure of the ratio ILO >2/1 2 relative to >1/0 silicosis. The number of missed cases per 1000 exposed persons 3 and number needed to screen to identify 1 extra case of silicosis in three scenarios 4 of low silicosis prevalence (5% or 50 cases per 1000), as defined by CXR ILO >1/0, 5 medium silicosis prevalence (15% or 150 cases per 1000) and high silicosis 6 prevalence (30% or 300 cases per 1000), and under the assumptions of fixed and 7 relative sensitivity of CXR, compared to HRCT. For the scenario of relative 8 sensitivity, we have further assumed either less common severe disease (20% of 9 cases are > ILO 2/1 relative to >1/0 silicosis) and more common severe disease 10 (40% of cases are > ILO 2/1 relative to >1/0 silicosis). 11 12 Low prevalence (5%) Medium prevalence (15%) High prevalence (30%) Missed cases per 1000 persons NNS to detect 1 extra case Missed cases per 1000 persons NNS to detect 1 extra case Missed cases per 1000 persons NNS to detect 1 extra case Fixed CXR sensitivity 16 62 47 21 95 11 Relative CXR sensitivity Less severe disease (20% > ILO 2/1) 79 13 236 4 471 2 More severe disease (40% > ILO 2/1) 31 32 93 11 185 5 13 14 15 16 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 13

1 2 Despite concerns regarding the diagnostic accuracy of CXR for silicosis, no 3 systematic review or meta-analysis has been performed to date. We found that, for 4 the diagnosis of ≥ 1/0 silicosis, CXR had a moderate pooled sensitivity compared to 5 HRCT (0.76; 95% CI 0.63, 0.86) and poor sensitivity compared to autopsy (0.50; 6 95% CI 0.45, 0.55). Specificity was high for both autopsy and HRCT. Heterogeneity 7 was high for HRCT (I2 = 84%) while two large autopsy studies reported identical 8 sensitivity values. 9 10 The sensitivity of CXR in detecting silicosis increased across autopsy, HRCT and 11 CT, reflecting the improved resolution and hence diagnostic accuracy of autopsy 12 over HRCT, and HRCT over CT. That the sensitivity of CXR compared to CT 13 approached 1 suggests that plain CT, as applied at the time of the included studies 14 (1986-1993), had little benefit for the screening detection of silicosis beyond 15 standard CXR. Older plain CT technology has been superseded by HRCT or 16 reconstructed MDCT images, thus limiting the generalisability of our CT results to 17 current CT imaging. Only one, small study from 1991 directly compared CT and 18 HRCT; although this reported equal sensitivity, the authors commented on improved 19 clarity and firmer interpretation with HRCT compared to CT [17]. Technological 20 improvements over time in both CXR and reference tests may influence results. 21 Older HRCT studies tended to have reduced sensitivity compared to more recent 22 studies (Figure 3). This suggests greater improvement in CXR accuracy over time 23 relative to HRCT, although these studies were also often smaller. 24 25 We observed increasing CXR sensitivity at higher cut-off grades of reference test 26 silicosis. This supports the intuitive and previously reported understanding that 27 severe silicosis is more readily diagnosed on CXR[8,9]. We then found that CXR 28 sensitivity increased by 1.15% (95% CI 0.44, 1.86) for every 1% increase in the ratio 29 of CXR ILO >2/1 relative to ≥ 1/0 silicosis, although heterogeneity was considerable 30 (I2 = 86%). 31 32 Based on our findings, it is plausible that both silicosis severity and reference test 33 type influence the diagnostic accuracy of CXR for silicosis. In the case of HRCT, it 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 14 may therefore be considered that the true sensitivity and specificity of CXR lies 1 between the assumptions of fixed and relative sensitivity. It is therefore important to 2 note that, regardless of assumption, a clinically meaningful number of cases were 3 missed (Table 1 and supplementary figure 7). Furthermore, higher silicosis 4 prevalence and greater proportions of severe disease led to higher numbers of 5 missed cases and a lower NNS. This suggests HRCT may have the greatest impact 6 among high prevalence populations with more severe disease. Conversely, although 7 the absolute impact of HRCT may be reduced among lower prevalence groups, if the 8 relative sensitivity scenario is considered probable, the proportional benefit from the 9 earlier diagnosis of silicosis is greater, particularly in the plausible case of less 10 frequent severe disease. 11 12 From a pragmatic perspective, our results would suggest a low threshold for HRCT 13 screening among groups or individuals considered to be at a higher risk of silicosis. 14 From the patient and physician perspective, although no established treatment for 15 silicosis exists, the use of HRCT may allow earlier diagnosis of silicosis, leading to 16 more informed work choices. For employers, industrial hygienists and regulators, 17 more accurate, earlier-stage silicosis prevalence data may prompt pro-active 18 exposure testing, adjustments to work practices and legislation. 19 20 Given the dose-response relationship between silica and lung cancer, and often high 21 rates of smoking among silica-exposed workers[3,35], our analysis represents the 22 minimum potential benefit of HRCT screening. Practical considerations of HRCT 23 screening include accessibility, cost-effectiveness, radiation dose and risks of 24 subsequent investigations, such as lung biopsy of high-risk nodules. Correctly risk 25 stratifying nodules among silica-exposed populations, particularly if TB is common, 26 presents an unexplored challenge. Low-dose CT screening with high resolution 27 reconstruction, such as implemented for lung cancer screening and silicosis 28 screening in Western Australia, is possible[11]. An alternative approach may involve 29 a combination of computer-aided detection for CXR, trained with HRCT as a 30

standard. If successful, this approach would have global impacts for the 31 diagnosis and classification of silicosis; our results can help design studies. 32 33 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 15 Individual study methods may influence sensitivity or specificity. Recruitment from 1 largely pre-screened silicosis populations, in whom prevalences of silicosis are 2 notably high (range 16-99%), may falsely raise sensitivity and reduce specificity. 3 Differences in test reader experience or reading and consensus methods may 4 increase between study variance. However, restriction to studies without high or 5 unclear risk of bias in these domains did not alter our findings. Future studies should 6 aim for an unselected population and a standardised reading and consensus 7 approach. Comorbidities including emphysema or tuberculosis did not appear to 8 reduce sensitivity or specificity[32,33]. 9 10 Important limitations of our methods exist. Whilst we applied our risk of bias tool 11 according to the guidance[13] and have attempted to do so transparently (including 12 publication of our raw data), we felt the QUADAS-2 may over-estimate risk of bias in 13 some cases. For example, the preselection of a sample (such as in Hoy et al (29)) 14

in a penalisation of both Patient selection and Flow and Timing, possibly 15 over-stating the true risk of bias. Potentially exacerbated by our English-language 16 restriction, our data is skewed to developed economies with relatively low 17 exposures[3] . When no ILO > 2/1 cases were present modelled sensitivity was 16%. 18 This may represent inaccuracy outside our data range as our lowest ILO >2/1 19 relative to >1/0 percentage was 38%. However, one study of only ILO 0 and 0/1 20 CXRs in miners found 57/339 (17%) had silicosis on HRCT suggesting this low 21 sensitivity is plausible[36]. Wide confidence intervals in our meta-regression 22 represent data sparsity and heterogeneity and lend caution to our subsequent 23 missed cases and NNS calculations. Whilst we maintain these estimates are useful, 24 they should be viewed as illustrative estimates of the impact of changes in 25 parameters – such as proportion of severe disease – rather than exact figures. Our 26 sensitivity analysis (using coefficients from the prevalence of ILO >2/1 exposure 27 model) demonstrates similar values, particularly at higher prevalences of silicosis. At 28 lower ranges, the differences between the missed case and NNS estimates of our 29 primary and sensitivity analysis suggests further caution regarding figures outside of 30 our data range. Whilst severity of disease is likely a major determinant of sensitivity, 31 other important factors that we did investigate that may determine CXR sensitivity 32 include the quality of images and experience of the reader and who hired them [37]. 33 34 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 16 Compared to HRCT, CXR has moderate sensitivity and high specificity for silicosis 1 diagnosis. Both the type of reference test and the proportion of severe disease 2 plausibly influence CXR sensitivity. Regardless, using CXR for silicosis screening 3

in missed cases of silicosis and – in most scenarios – a relatively low number 4 of participants are needed to be screened with HRCT to detect a single extra case of 5 silicosis. 6 7 Conflict of interest statement: No conflicts of interest 8 9 Funding support: PH - Medical Research Council (MR/W024861/1). ML is partially 10 supported by the Academy of Medical Sciences Professorship (APR7\1005). No 11 other financial disclosures 12 13 Notation of prior abstract publication/presentation: Accepted for presentation by 14 Ashwin Durairaj at the British Thoracic Society Winter meeting, London, United 15 Kingdom, 27-29th November 2024 16 17

18 PH – Guarantor: full access to all the data in the study and takes responsibility for 19 the integrity of the data and the accuracy of the data analysis, including and 20 especially any adverse effects, Writing – Original draft preparation, 21 Conceptualisation, Methodology, Project Administration, Data collection, Data 22 Curation, Visualisation, Formal Analysis 23 AD – Original draft preparation, Conceptualisation, Methodology, Project 24 Administration, Data collection, Data Curation, Visualisation, Formal Analysis 25 ML - Conceptualisation, Formal Analysis, Visualisation, Writing – Review & Editing 26 JF – Conceptualisation, Project Administration, Supervision, Writing – Review & 27 Editing 28 29 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 17 Figure 1. Directed acyclic diagrams of the conceptual relationship between 1 cumulative silica exposure and silicosis. Two counterfactual relationships are 2 represented. A. The relationship between cumulative silica exposure and radiological 3 silicosis is fixed – sensitivity does not vary according to disease severity. B. The 4 relationship between cumulative silica exposure and identification on radiological test 5 is mediated and thus relative to the severity of silicosis – sensitivity varies relative to 6 disease severity. 7 8 Figure 2. PRISMA Diagram 9 10 Figure 3. Forest plots of sensitivity and specificity from meta-analysis of chest X-ray 11 (CXR), at an ILO >1/0 cut-off, for the diagnosis of silicosis. 12 13 A. Describes the sensitivity (left) and specificity (right) of CXR compared to high-14 resolution computed tomography (HRCT). 15 16 B. Describes the sensitivity (left) and specificity (right) of CXR compared to autopsy 17 18 Random effects model uses a generalised linear mixed model. Heterogeneity is 19 determined through the I2 statistic and Chi-squared (χ ²) test. 20 21 Abbreviations: C.I, confidence interval. 22 23 Figure 4. A mixed-effects linear meta-regression model showing the association 24 between the sensitivity of CXR at ILO category ≥ 1 and the ratio ILO >2/1 relative to 25 >1/0 silicosis across 7 studies. Each circle represents a study, with the diameter 26 proportional to the study's size. The circle is coloured depending on reference 27 standard (Red = autopsy, Blue = computed tomography and Green = high resolution 28 computed tomography). The solid line represents the predicted sensitivity based on 29 the ratio ILO >2/1 relative to >1/0 silicosis. The model explains a moderate 30 proportion of the heterogeneity in CXR sensitivity between studies (R² = 63%, I2 = 31 86%). 32 33 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 18

1 1. Feary J, Devaraj A, Burton M, Chua F, Coker RK, Datta A, et al. Artificial stone 2 silicosis: a UK case series. Thorax. 2024 Sep 18;79(10):979–81. 3 2. Fazio JC, Gandhi SA, Flattery J, Heinzerling A, Kamangar N, Afif N, et al. 4 Silicosis Among Immigrant Engineered Stone (Quartz) Countertop Fabrication 5 Workers in California. JAMA Intern Med. 2023 Sep 1;183(9):991–8. 6 3. Howlett P, Mousa H, Said B, Mbuya A, Kon OM, Mpagama S, et al. Silicosis, 7 tuberculosis and silica exposure among artisanal and small-scale miners: A 8 systematic review and modelling paper. PLOS Glob Public Health. 9 2023;3(9):e0002085. 10 4. Chen S, Liu M, Xie F. Global and national burden and trends of mortality and 11 disability-adjusted life years for silicosis, from 1990 to 2019: results from the 12 Global Burden of Disease study 2019. BMC Pulm Med. 2022 Jun 21;22(1):240. 13 5. International Labour Office, editor. Guidelines for the use of the ILO international 14 classification of radiographs of pneumoconioses. Revised edition 2011. 15 Geneva: International Labour Office; 2011. 48 p. (Occupational safety and 16 health series). 17 6. OSHA. Occupational Exposure to Respirable Crystalline Silica. Final rule. 18 [Internet]. Federal Register (USA); 2016 p. 16285–890. Report No.: 81(58). 19 Available from: https://www.federalregister.gov/documents/2016/03/25/2016-20 04800/occupational-exposure-to-respirable-crystalline-silica 21 7. Health and Safety Executive (UK). Health surveillance for those exposed to 22 respirable crystalline silica [Internet]. [cited 2024 Aug 22]. Available from: 23 https://www.hse.gov.uk/pubns/priced/healthsurveillance.pdf 24 8. Hnizdo E, Murray J, Sluis-Cremer GK, Thomas RG. Correlation between 25 radiological and pathological diagnosis of silicosis: An autopsy population based 26 study. Am J Ind Med [Internet]. 1993 Oct [cited 2022 Oct 25];24(4):427–45. 27 Available from: https://onlinelibrary.wiley.com/doi/10.1002/ajim.4700240408 28 9. Ehrlich R, Murray J, Rees D. Subradiological silicosis. Am J Ind Med [Internet]. 29 2018 Nov [cited 2022 Oct 25];61(11):877–85. Available from: 30 https://onlinelibrary.wiley.com/doi/10.1002/ajim.22909 31 10. Hoy RF, Jones C, Newbigin K, Abramson MJ, Barnes H, Dimitriadis C, et al. 32 Chest x-ray has low sensitivity to detect silicosis in artificial stone benchtop 33 industry workers. Respirology. 2024 Sep;29(9):785–94. 34 11. Silicosis Position Statement | RANZCR [Internet]. [cited 2024 Oct 16]. Available 35 from: https://www.ranzcr.com/college/document-library/silicosis-position-36 statement 37 12. All-Party Parliamentary Group for Respiratory Health. APPG Report: Improving 38 Silicosis outcomes in the UK [Internet]. UK Government; 2023 Jan. Available 39 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 19 from: https://www.appg-respiratory.co.uk/sites/appg/files/2023-01/APPG-1 Silicosis-Report-01-23.pdf 2 13. Whiting PF, Rutjes AWS, Westwood ME, Mallett S, Deeks JJ, Reitsma JB, et al. 3 QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy 4 studies. Ann Intern Med. 2011 Oct 18;155(8):529–36. 5 14. Hodous TK, Chen RA, Kinsley KB, Liu XT, McLaughlin JK, Chen JQ, et al. A 6 comparison of pneumoconiosis interpretation between Chinese and American 7 readers and classifications. J Tongji Med Univ. 1991;11(4):225–9. 8 15. Tamura T, Suganuma N, Hering KG, Vehmas T, Itoh H, Akira M, et al. 9 Relationships (I) of International Classification of High-resolution Computed 10 Tomography for Occupational and Environmental Respiratory Diseases with the 11 ILO International Classification of Radiographs of Pneumoconioses for 12 parenchymal abnormalities. Ind Health. 2015;53(3):260–70. 13 16. Harrer M, Cuijpers P, Furukawa TA, Ebert DD. Doing Meta-Analysis With R: A 14 Hands-On Guide [Internet]. 1st ed. Boca Raton, FL and London: Chapman & 15 Hall/CRC Press; 2021. Available from: https://www.routledge.com/Doing-Meta-16 Analysis-with-R-A-Hands-On-Guide/Harrer-Cuijpers-Furukawa-17 Ebert/p/book/9780367610074 18 17. Bégin R, Ostiguy G, Fillion R, Colman N. Computed Tomography Scan in the 19 Early Detection of Silicosis. Am Rev Respir Dis [Internet]. 1991 Sep [cited 2024 20 Sep 3];144(3_pt_1):697–705. Available from: 21 http://www.atsjournals.org/doi/abs/10.1164/ajrccm/144.3_Pt_1.697 22 18. Grenier P, Valeyre D, Cluzel P, Brauner MW, Lenoir S, Chastang C. Chronic 23 diffuse interstitial lung disease: diagnostic value of chest radiography and high-24 resolution CT. Radiology. 1991 Apr;179(1):123–32. 25 19. Talini D, Paggiaro PL, Falaschi F, Battolla L, Carrara M, Petrozzino M, et al. 26 Chest radiography and high resolution computed tomography in the evaluation 27 of workers exposed to silica dust: relation with functional findings. Occup 28 Environ Med. 1995 Apr;52(4):262–7. 29 20. Antao VCDS, Pinheiro GA, Terra-Filho M, Kavakama J, Muller NL. High-30 Resolution CT in Silicosis: Correlation With Radiographic Findings and 31 Functional Impairment. Journal of Computer Assisted Tomography [Internet]. 32 2005 May [cited 2024 Sep 9];29(3):350–6. Available from: 33 http://journals.lww.com/00004728-200505000-00015 34 21. Murgia N, Muzi G, Dell’Omo M, Sallese D, Ciccotosto C, Rossi M, et al. An old 35 threat in a new setting: High prevalence of silicosis among jewelry workers. Am 36 J Ind Med. 2007 Aug;50(8):577–83. 37 22. Sun J, Weng D, Jin C, Yan B, Xu G, Jin B, et al. The Value of High Resolution 38 Computed Tomography in the Diagnostics of Small Opacities and 39 Complications of Silicosis in Mine Machinery Manufacturing Workers, 40 Compared to Radiography. Jrnl of Occup Health [Internet]. 2008 Sep [cited 41 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 20 2023 Jul 11];50(5):400–5. Available from: 1 https://onlinelibrary.wiley.com/doi/abs/10.1539/joh.L8015 2 23. Lopes AJ, Mogami R, Capone D, Tessarollo B, de Melo PL, Jansen JM. High-3 resolution computed tomography in silicosis: correlation with chest radiography 4 and pulmonary function tests. J Bras Pneumol. 2008 May;34(5):264–72. 5 24. Meijer E, Tjoe Nij E, Kraus T, van der Zee JS, van Delden O, van Leeuwen M, 6 et al. Pneumoconiosis and emphysema in construction workers: results of 7 HRCT and lung function findings. Occup Environ Med. 2011 Jul;68(7):542–6. 8 25. Berk S, Dogan DO, Gumus C, Akkurt I. Relationship between radiological (X-9 ray/HRCT), spirometric and clinical findings in dental technicians’ 10 pneumoconiosis. Clin Respir J. 2016 Jan;10(1):67–73. 11 26. Takahashi M, Nitta N, Kishimoto T, Ohtsuka Y, Honda S, Ashizawa K. 12 Computed tomography findings of arc-welders’ pneumoconiosis: Comparison 13 with silicosis. Eur J Radiol. 2018 Oct;107:98–104. 14 27. Ş ener MU, Ş imş ek C, Özkara Ş , Evran H, Bursali İ , Gökçek A. Comparison of 15 the International Classification of High-resolution Computed Tomography for 16 occupational and environmental respiratory diseases with the International 17 Labor Organization International Classification of Radiographs of 18 Pneumoconiosis. Industrial Health [Internet]. 2018 Oct 19 [cited 2024 Oct 19 17];57(4):495. Available from: 20 https://pmc.ncbi.nlm.nih.gov/articles/PMC6685792/ 21 28. León-Jiménez A, Hidalgo-Molina A, Conde-Sánchez MÁ, Pérez-Alonso A, 22 Morales-Morales JM, García-Gámez EM, et al. Artificial Stone Silicosis: Rapid 23 Progression Following Exposure Cessation. Chest. 2020 Sep;158(3):1060–8. 24 29. Hoy RF, Dimitriadis C, Abramson M, Glass DC, Gwini S, Hore-Lacy F, et al. 25 Prevalence and risk factors for silicosis among a large cohort of stone benchtop 26 industry workers. Occup Environ Med [Internet]. 2023 Aug [cited 2024 Sep 27 3];80(8):439–46. Available from: 28 https://oem.bmj.com/lookup/doi/10.1136/oemed-2023-108892 29 30. Crawford JA, Sanyal S, Burnett BR, Wiesenfeld SL, Abraham JL. Accelerated 30 silicosis in sandblasters: Pathology, mineralogy, and clinical correlates. Am J 31 Ind Med. 2024 Mar;67(3):179–99. 32 31. Bergin C, Muller N, Vedal S, Chan-Yeung M. CT in silicosis: correlation with 33 plain films and pulmonary function tests. American Journal of Roentgenology 34 [Internet]. 1986 Mar 1 [cited 2024 Sep 3];146(3):477–83. Available from: 35 https://www.ajronline.org/doi/10.2214/ajr.146.3.477 36 32. Cowie RL, Hay M, Thomas RG. Association of silicosis, lung dysfunction, and 37 emphysema in gold miners. Thorax [Internet]. 1993 Jul 1 [cited 2024 Sep 38 3];48(7):746–9. Available from: 39 https://thorax.bmj.com/lookup/doi/10.1136/thx.48.7.746 40 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 21 33. Corbett EL, Murray J, Churchyard GJ, Herselman PC, Clayton TC, De Cock 1 KM, et al. Use of Miniradiographs to Detect Silicosis: Comparison of 2 Radiological with Autopsy Findings. Am J Respir Crit Care Med [Internet]. 1999 3 Dec 1 [cited 2023 Apr 14];160(6):2012–7. Available from: 4 https://www.atsjournals.org/doi/10.1164/ajrccm.160.6.9903040 5 34. Bégin R, Bergeron D, Samson L, Boctor M, Cantin A. CT assessment of 6 silicosis in exposed workers. AJR Am J Roentgenol. 1987 Mar;148(3):509–14. 7 35. Wang D, Yang M, Liu Y, Ma J, Shi T, Chen W. Association of Silica Dust 8 Exposure and Cigarette Smoking With Mortality Among Mine and Pottery 9 Workers in China. JAMA Netw Open [Internet]. 2020 Apr 14 [cited 2023 Feb 10 25];3(4):e202787. Available from: 11 https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2764378 12 36. Carneiro APS, Da Silva LL, Silva FDCL, Hering KG, Algranti E. Volume-based 13 tomography for the diagnosis of incipient silicosis in former gold miners. Occup 14 Environ Med [Internet]. 2022 Jun [cited 2024 Sep 12];79(6):427–32. Available 15 from: https://oem.bmj.com/lookup/doi/10.1136/oemed-2021-107922 16 37. Akgün M, Ozmen I, Ozari Yildirim E, Tuzun B, Nur Toreyin Z, Kayinova A, et al. 17 Pitfalls of using the ILO classification for silicosis compensation claims. 18 Occupational Medicine [Internet]. 2022 Jul 1 [cited 2025 Mar 19];72(6):372–7. 19 Available from: https://doi.org/10.1093/occmed/kqac010 20 21 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint 22 . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint . CC-BY 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted May 29, 2025. ; https://doi.org/10.1101/2025.05.28.25328086doi: medRxiv preprint