Computer-Aided Analysis of Photographed Chest X-Ray Films Performs Well Compared With Trained Radiologists

Abstract

Objective

To assess whether computer-aided detection (CAD) chest X-ray (CXR) software may aid physicians in low-resource, high tuberculosis (TB) endemic settings where radiologists are scarce.

Patients and Methods

A retrospective pilot study was conducted on CXR films taken between January 1, 2017, and March 30, 2018, in Guinea-Bissau and Ethiopia to compare the interpretation of CXRs regarding pulmonary TB (PTB) by CAD (qXR; Qure.ai) with that of 2 experienced Ethiopian radiologists (A and B). To improve the applicability of this method in low-resource settings, an analysis was performed on images of CXRs taken by mobile phones. Two reference standards were applied: final PTB diagnosis by clinical or laboratory findings (ie, Xpert MTB/RIF [Xpert]-confirmed PTB).

Results

We included 498 CXRs from patients seeking help for TB indicative symptoms. Radiologist A identified 50, radiologist B identified 99, and the software identified 81 as indicative of TB. The overall area under the curve for the receiver-operating characteristic curve of the software was 0.84 for Xpert-confirmed cases. At the prechosen cutoff value of 0.5, the sensitivity of CAD CXR was 76.5%, and the specificity was 85.9%. Radiologist A’s assessments were 64.7% sensitive and 91.9% specific, whereas radiologist B’s assessments were 76.5% sensitive and 82.3% specific for Xpert-confirmed cases. The agreement regarding TB-related findings between the radiologists combined (κ=0.45) and each radiologist and the software (κ=0.56) was moderate.

Conclusion

Our study revealed that CAD CXR performs comparably with experienced radiologists when it is applied to CXR films, photographed by mobile phones and a digital camera with similar sensor resolutions.

Trial registration

PACTR201611001838365.

Tuberculosis (TB) remains a major threat to global health with an estimated 10.8 million people experiencing TB in 2023 and 1.09 million deaths due to the disease. Although the World Health Organization (WHO) aims at a 90% reduction in TB incidence by 2035,¹ case detection rates are still insufficient to achieve this goal.²

With an estimated 2.7 million undiagnosed cases in 2023,¹ the need to develop novel strategies and technologies aimed at improving case detection in low-resource, high-incidence settings is urgent. In most low-resource settings, case finding is predominantly passive and patients with presumed TB must be identified among adults presenting at primary health care. Moreover, many people with TB are missed at first contact because TB is not suspected.3, 4, 5

Sputum samples and smear microscopy have been the cornerstone of TB diagnosis in highly endemic areas for decades, albeit with substantially varying sensitivities, ranging from 20% to 50%.⁶ The advent of PCR-based point-of-care solutions such as Xpert MTB/RIF (Xpert) has increased sensitivity to 60% to 70%; however, this approach leaves ample room for improvement.^7,8

A recent meta-analysis suggested that chest x-ray (CXR) may substantially improve TB detection rates and called for the evaluation of computer-aided detection (CAD) platforms for TB screening.⁹ Chest x-ray remains a diagnostic tool in smear-negative patients with presumed TB in low-resource areas; however, there is still a need for further studies on costs, implementation aspects, the integration of CAD into diagnostic algorithms, and performance in key groups.¹⁰ The WHO conditionally recommends the use of CAD CXRs in the population where TB screening is needed.¹¹ Computer-aided detection software based on artificial intelligence (AI) algorithms has been shown to improve the sensitivity of detecting microbiologically confirmed TB from 50% to 60% (with a specificity of 80%) by experienced radiologists to 66% to 76%.¹²

Thus, CAD CXR may represent a useful add-on, following initial microbiological testing as recommended by the WHO, to improve the detection and diagnosis of TB in areas where experienced radiologists are scarce, and microbiological methods fail or are unavailable, and the diagnosis is based on a clinical assessment. However, previous studies have focused on the analysis of digital CXRs, limiting applicability in low-resource settings where CXRs are frequently available only in analog form.

In 2019, there were 6 conformité européenne–marked CAD products for TB: CAD4TB, qXR, Inferread dr chest, JViewer-X, Lunit insight CXR, and AXIR.¹³ Previous studies have evaluated the diagnostic accuracy of qXR against human readers using Xpert MTB/RIF¹² as a reference standard. In a study conducted in India, qXR showed moderate specificity and sensitivity for detecting pulmonary TB.¹² Another study showed higher specificity than that of radiologists, while maintaining similar sensitivity.¹⁴ An evaluation of multiple AI algorithms intended for detecting radiological signs of TB in CXRs using data from patients in a high TB-burden setting in Bangladesh reported that qXR has an area under the curve (AUC) of 0.9081.¹⁵ Similarly, an independent evaluation of 12 different AI algorithms for TB detection conducted in Vietnam revealed a receiver-operating characteristic (ROC) AUC for qXR of approximately 0.82.¹⁶

In recent years, those CAD products have been evaluated with regards to their compatibility with smartphone captured pictures of CXR films, acknowledging that many high endemic settings still use traditional X-ray films.¹⁷ Those studies have shown that it indeed is possible to analyze smartphone images of traditional X-ray but that it needs careful recalibration.¹⁸

The present pilot study aimed to assess the diagnostic value of qXR for TB using analog CXR films photographed by mobile phones or digital cameras. The analysis used CXRs from patients seeking help for TB indicative symptoms at public health centers in low-resource, high-endemic countries of Ethiopia and Guinea-Bissau and compared the results with assessments by 2 Ethiopian radiologists. To our knowledge, there is little evidence of the performance of CAD CXRs on analog images. Moreover, successful mobile phone–based analog radiographic image analysis could represent an easily implementable tool for screening and reduce diagnostic barriers for people living in low-resource settings. This study therefore provides vital evidence for the implementation of CAD CXRs in low-resource settings with limited access to radiologists and digital CXRs.

Patients and Methods

Design, Study Population, and TB Diagnosis

This retrospective diagnostic pilot study included a set of 498 digitally photographed chemically processed posterior-anterior (PA) CXR films of 322 Ethiopian and 176 Bissau-Guinean patients included in a cluster randomized clinical trial conducted in Ethiopia and Guinea-Bissau from January 2017 until the end of March 2018^19,20 (Figure 1). The patients included were adults seeking help for symptoms indicative of pulmonary TB (PTB) at primary health care centers. They were part of a follow-up group seen 2 weeks after inclusion where initial sputum analysis was not performed or was negative. The CXR and Xpert analyses were performed within the same week prior to the clinical assessment and were only carried out for patients who had persistent symptoms at the 2-week follow-up. Not all patients were able to produce a sputum sample. For those who were unable to do so, the diagnosis was based on clinical assessment and CXR.

Figure 1.

Study flow diagram, by diagnosis. CAD, computer-aided detection; CXR, chest x-ray; TB, tuberculosis.

Clinical PTB was diagnosed by experienced physicians based on clinical signs and symptoms, CXR findings and a trial of antibiotics.²⁰ This diagnostic standard was included because it is a common method for the diagnosis of TB in low-resource settings.²¹ The clinicians assessing the patients were blinded to the radiologists’ and CAD assessments because they were performed after the study conclusion.

The incidence of TB in Ethiopia is estimated to 146 of 100,000, whereas it is estimated to 361 of 100,000 in Guinea-Bissau.²² In both settings, more males are diagnosed with TB disease and the male:female ratio for Ethiopia is 2.17 and 1.37 for Guinea-Bissau.²³

Disease Severity Measured by the TBscore

To quantify disease severity and qualify the clinical diagnosis of TB further, we used the Bandim TBscore. It has previously been assessed regarding disease severity and diagnostic capability and consists of TB-related signs and symptoms.^20,24

The CAD CXR

The software qXR was developed by qure.ai (Mumbai, India) as an AI-based CAD for analyzing PA CXRs for abnormalities.²⁵ The AI algorithm was trained using a multihospital training data set of 4 million CXRs and their corresponding radiology reports to identify abnormal X-rays and specific abnormalities.

The software takes the pixel values of first color channel as its input for further downstream processing. For example, if the color channels are in RGB, then the pixel values of R channel will be taken.

The qXR software assigns a score between 0 and 1 to indicate the presence of a certain finding. This study used a general qXR threshold of 0.5 for PTB, which is considered the default recommended for most settings.²⁶ The qXR software also outputs a pixel map, indicating the location of the target abnormality (Supplemental Figure 1, available online at https://www.mcpdigitalhealth.org/). Each radiological finding also has a certain predefined threshold at which it is deemed positive. In the present study, qXR, version 3.2, was used to analyze the CXRs.

The Radiologists

Radiologist A is an assistant professor of radiology at the College of Medical and Health Sciences, University of Gondar (CMHS UOG) and obtained his medical degree at CMHS UOG and his radiology specialty certificate at Addis Ababa University. In addition, he has served in an academic position and as a consultant radiologist at Gondar University Hospital (GUH). He has served as the head of the Department of Radiology at GUH for 2 years. He has 10 years of experience in the field of radiology.

Radiologist B is an associate professor of radiology at the Department of Radiology at CMHS UOG. He has worked as a consultant radiologist at the GUH since 2010. He obtained his medical degree and specialty certificate in radiology at Addis Ababa University. He currently serves as head of department at the Department of Radiology at CMHS UOG. He has 14 years of experience in the field of radiology.

We also analyzed the pooled radiologists’ assessments against CAD CXRs. We pooled the radiologists’ assessments by coding a finding as positive if 1 of the radiologists had identified it on the CXR. The radiologists and AI crew were blinded to all clinical information and microbiological test results.

Assessment of CXR

The chemically processed PA CXR films were photographed using different commonly available digital or mobile phone cameras available in the different study settings (Sony Cyber-shot [20.1 MP]; Samsung Galaxy A3 [8.0 MP]; and Canon EOS 70D [20.2 MP]; resolution varying from 72 to 350 dpi), with the CXRs on a light box in daylight, switching off the flashlights using a guide provided by qXR²⁷ (Supplemental Figure 2, available online at https://www.mcpdigitalhealth.org/), anonymized, numbered in a random order, and sent digitally to qure.ai and the radiologists. The files were stored as jpg. The compression level for Canon EOS 70D set at large/fine, whereas it was standard for the Sony Cyber-shot. Thus, both the radiologists and CAD CXR analyzed digital pictures of chemically processed CXR films.

All cameras did supply images from patients diagnosed with TB either by Xpert or clinically (Tables 1 and 2). Sample input images from each camera can be seen in Supplemental Figure 1.

Table 1.

Baseline Characteristics, by TB Diagnosis

Characteristic	All (N=498)	PTB diagnosis (n=58), n (12%)	No PTB diagnosis (n=440), n (88%)	P (TB vs not TB)
Female	259 (52)	23 (40)	236 (54)	.05
Age, y	38 (23-50)	43 (25-62)	37 (22-50)	.01
TBscore	4·4 (3-6)	5.9 (4-7)	4.2 (3-6)	<.001
Previous TBa	54 (11)	12 (21)	42 (10)	.06
HIV-infectedb	28 (6)	3 (5)	25 (6)	.87
Camera type				.66
Samsung	301 (60)	38 (66)	263 (60)
Canon	125 (25)	13 (22)	112 (26)
Sony	72 (15)	7 (12)	65 (14)

Values are n (%) or mean (IQR).

Missing: ^a 18 (4%); ^b 204 (41%).

Table 2.

Characteristics for TB Diagnosed, By Diagnosis

Characteristic	All (N=58)	Clinical (n=41l; 72%)	GenXpert (n=17; 28%)	P (GenXpert vs clinical)
Female	23 (40)	18 (44)	5 (29)	.08
Age	43 (25-62)	45 (35-63)	34 (25-40)	.99
TBscore	5.8 (4-7)	6.1 (5-7)	5.4 (3-8)	.86
Previous TBa	12 (21)	8 (20)	4 (24)	.27
HIV: infected	3 (5)	2 (5)	1 (6)	.88
Camera type				.04
Samsung	38 (66)	31 (76)	7 (41)
Canon	13 (22)	7 (17)	6 (35)
Sony	7 (12)	3 (7)	4 (24)

Values are n (%) or mean (IQR).

Missing: ^a1.

Classification of Findings

Chest x-rays were classified as abnormal if 1 of the following findings was identified: consolidation, cavitation, fibrosis, nodule, blunted cardiopulmonary angle, pleural effusion, hilar lymphadenopathy, or tracheal shift.

Statistical Analyses

All data analyses were performed using Stata version 11 (Stata Corporation). Kappa coefficients, and associated 95% CIs were used to investigate the interrater reliability of radiologists and AI. The following scale was used for the interpretation of κ coefficients: <0, poor; 0 to 0.20, slight; 0.21 to 0.40, fair; 0.41 to 0.60, moderate; 0.61 to 0.80, substantial; and 0.81 to 1.00, almost perfect.²⁸ To assess the overall accuracy of the AI, ROC analysis was used to determine the AUC using the “roctab” command in Stata 11, which provides a nonparametric estimation of the ROC curve and Bamber and Hanley CIs for the AUC of the ROC curve. The sensitivity and specificity were calculated. A Venn diagram was constructed using the user written command venndiag²⁹ in Stata software.

The same accuracy measures were also used to compare the radiologist’s assessment with that of the CAD CXR and were plotted together with the ROC of CAD CXR. The sensitivity, specificity, positive predictive value, and negative predictive value were calculated, and microbiological confirmation (by Xpert) and/or clinical diagnosis were used as the reference standards.

Results

The patients contributing the 498 CXRs to the analysis had a mean age of 38 years (IQR, 23-50 years), 259 (52%) were females, and 28 (6%) were infected with HIV. Fifty-four (11%) reported having had TB previously (Table 1). In total, diagnosis of TB was made in 58 (12%) patients (Table 2) of whom 17 (28%) were confirmed to have a positive Xpert, and 40 (69%) were from Ethiopia.

The percentage of females was lower among patients diagnosed with TB (40%) than that among patients not diagnosed with TB (54%; P=.05). The TBscore was higher among patients diagnosed with TB (5.9; IQR, 4-7) than that among patients not diagnosed with TB (4.2; IQR, 3-6; P<.001) (Table 1).

qXR Interpretation—GX or Clinically Confirmed Cases

The ROC curves computed from the scores assigned by the CAD CXR readings are presented in Figure 2. The AUC of CAD CXR was 0.78 (95% CI, 0.78-0.85) for overall TB diagnosis, 0.84 (95% CI, 0.73-0.96) for Xpert-confirmed cases and 0.74 (95% CI, 0.65-0.82) for clinically diagnosed patients.

Figure 2.

Receiver-operating characteristics (ROC): (A) all TB (clinical and GX confirmed); (B) only GX confirmed; (C) only clinically diagnosed; star, radiologist A; triangle, radiologist B. AUC, area under the curve.

At the prespecified cutoff value of 0.5, CAD CXR sensitivity ranged from 46.3% for clinically diagnosed patients to 76.5% for Xpert-confirmed patients, whereas the specificity varied between 85.9% for Xpert-confirmed and 88.9% for Xpert or patients with clinically diagnosed PTB (Table 3).

Table 3.

Interpretation of Radiologist A, Radiologist B, Radiologists Combined, and qXR (at Prespecified Cutoff; GX Confirmed)

Evaluater	Cutoff score	TP	FP	FN	TN	Sensitivity	Specificity
All cases (GX or clinical)
Radiologist A	NA	24	26	34	414	41.4 (28.6-55.1)	94.1 (91.5-96.1)
Radiologist B	NA	32	66	26	374	55.2 (41.5-68.3)	85.0 (81.3-88.2)
Radiologists combined	NA	37	72	21	368	63.8 (50.1-76.0)	83.6 (79.8-87.0)
qXR	≥0.5	32	49	26	391	55.2 (41.5-68.3)	88.9 (85.5-91.6)
GX confirmed
Radiologist A	NA	11	39	6	442	64.7 (38.3-85.8)	91.9 (89.1-94.2)
Radiologist B	NA	13	85	4	396	76.5 (50.1-93.2)	82.3 (78.6-85.6)
Radiologists combined	NA	14	95	3	386	82.4 (56.6-96.2)	80.2 (76.4-83.7)
qXR	≥0.5	13	68	4	413	76.5 (50.1-93.2)	85.9 (82.4-88.9)
Clinical
Radiologist A	NA	13	37	28	420	31.7 (18.1-48.1)	91.9 (89.0-94.2)
Radiologist B	NA	19	79	22	378	46.3 (30.7-62.6)	82.5 (78.9-86.1)
Radiologists combined	NA	23	86	18	371	56.1 (39.7-71.5)	81.2 (77.3-84.7)
qXR	≥0.5	19	62	22	395	46.3 (30.7-62.6)	86.4 (82.9-89.4)

FN, false negative; FP, false positive; TN, true negative; TP, true positive.

Overall, the combined radiologists and CAD CXR agreed on 61 CXRs indicating TB, 13 of which were Xpert confirmed and 15 were clinically diagnosed, whereas 88 were not diagnosed with TB during the study period. The CAD CXR found 20 additional CXRs indicating TB, 4 of them clinically diagnosed, whereas the radiologists found 48 additional TB indicative CXRs, 8 of which were also clinically diagnosed, and 1 was Xpert confirmed (Supplemental Figure 3, available online at https://www.mcpdigitalhealth.org/). Examples of CXRs agreed and not agreed upon are shown in Supplemental Figure 1.

Comparison of qXR and Human Readers

Figure 2 displays the operating points for radiologists A and B together with the CAD CXR ROC for Xpert-confirmed, clinically diagnosed, and overall TB, respectively. Overall, the pooled radiologists’ assessments exhibited greater sensitivity than did the CAD CXR assessment and ranged from 56.1% (specificity, 81.2%) among clinically diagnosed to 82.4% (specificity, 80.2%) among Xpert-confirmed patients. Assessing each radiologist against the CAD CXR showed that radiologist A performed with less sensitivity but higher specificity, whereas radiologist B was as sensitive as the CAD CXR but performed with lower specificity (Table 3).

The radiologists’ agreement varied from slight (0.15; 95% CI, 0.04-0.26) for cardiomegaly to substantial (0.64; 95% CI, 0.45-0.82) for tracheal shift and was moderate (0.45; 95% CI, 0.35-0.55) for the overall identification of PTB (Supplemental Table 1, available online at https://www.mcpdigitalhealth.org/). Overall, the agreement between the radiologists’ pooled assessment and the CAD CXR regarding TB diagnosis based on CXR findings was moderate (0.56; 95% CI, 0.46-0.65).

The different prespecified findings (Supplemental Table 1) between the radiologists and CAD CXR ranged from fair (0.24; 95% CI, 0.06-0.42) for hilar lymphadenopathy to substantial (0.61; 95% CI, 0.48-0.74) and 0.61 (95% CI, 0.45-0.77) for nodule and tracheal shift, respectively.

Camera Type Distribution and Effect on Classification

Most of the pictures were taken by Samsung (60%), with no significant difference in percentage of TB cases per camera (Table 1). However, for radiologist B and CAD CXR, a proportionate higher share of TB classified CXRs were taken by Sony (32% and 29%, respectively, data not shown), which was significantly higher that the share of patients with TB with CXRs photographed by Sony (10%) (Table 1).

Discussion

This pilot study indicates that CAD CXR software applied to analog CXRs photographed by mobile phones or digital cameras may be as accurate in identifying TB as trained radiologists working in a high TB-endemic setting when assessing CXRs from patients seeking help for TB indicative symptoms at primary health care centers. Previous assessments of CAD software have been based on digital CXRs, and to the authors’ knowledge, the present study is the first to assess CAD applied to analog CXRs photographed by mobile phones or digital cameras. These findings thus have substantial implications for settings where TB is highly endemic and CXRs are still predominant analog.¹³

The use of CXR as a diagnostic tool for PTB has limitations. The sensitivity in previous studies ranged from 72% to 99% for any finding and 73% to 94% for findings suggestive of TB, whereas the specificity varied from 73% to 94% and 87% to 96%, respectively.^30,31 In the present study, both the radiologists and the CAD software were slightly less accurate at the prespecified cutoff value. This may most likely be because our study mainly included CXRs from smear-negative patients (ie, their CXRs may show patterns that are less specific to TB). Previous studies assessing the agreement between radiologists and CAD software found it to be fair,³² whereas we found it to be moderate between the radiologists and the radiologists and CAD software.

The operating points (ie, sensitivity and 1 −specificity) for both radiologists assessed in this study were very close to the ROC curve of the CAD software. This finding has been shown in previous studies assessing the qXR.^12,14,33

The present study included 498 individual CXRs from patients seeking help for TB indicative symptoms, 58 (12%) of which were diagnosed with TB including 17 (28%) of which were microbiologically confirmed by Xpert, which resembles data from previous studies.^19,20,34,35 These limited data are due to our use of CXRs from a previous case-detection study²⁰ and our focus on analog CXRs. The CXRs were taken during the diagnostic cascade of the study¹⁹ and thus belonged to individuals who were either smear negative initially or where sputum smear analysis was not considered necessary upon initial evaluation. The TBscore, included in this analysis to support the validity of the clinical diagnosed cases,^19,20,35 was high for the clinical diagnosed cases, thus reflecting a higher disease severity in this group where there was no microbiological confirmation. All individuals were symptomatic 2 weeks after the first contact with the health care provider where they had received antibiotic treatment. Thus, these patients represent an important group of patients, who are often overlooked in the cascade of care making up part of the considerable group of missed TB cases in highly endemic settings. Because the present study is retrospective, we were unable to confirm the cases diagnosed by CAD CXR only who did not receive a PTB diagnosis in the past study.

As seen in previous studies more women than men did seek help for TB indicative symptoms but more men were diagnosed with TB throughout the diagnostic process.^20,34,35 The clinical diagnosis was based on signs and symptoms after a trial of antibiotics. The assessment also included the CXR that was read by the physician assessing the patient. Thus, CXR is both part of reference and evaluation. Because TB in many parts of the world still is a clinical diagnosis that may or may not be partly based on CXRs read by the attending physician, we included it in our study to reflect current realities. The radiologist and CAD team did not receive information on signs and symptoms and thus based their assessments solely on the CXRs, whereas the clinician diagnosing the patients during the ongoing study based their diagnosis on both clinical examination and CXR. As shown by our study, the clinicians and radiologists are not in complete agreement regarding PTB diagnosis, and there is a subgroup of patients diagnosed with TB that were not identified by the clinicians as well as there is a group not identified by the CAD CXR/radiologists. The radiologists were not in complete agreement either, which is a well-known issue that reflects interrater reliability and depends on experience.³⁶ This strengthens the argument that there is a need and a place for CAD CXR in the diagnostic cascade.

Using a pooled radiologist assessment may introduce uncertainty in the assessment as there is no third assessor and thus no tie break for discordant results. Pooling the radiologists’ assessments led to a higher rate of abnormal CXRs in the pooled assessment possibly affecting specificity of the read.

The study did not confirm the diagnosis of PTB by mycobacterial culture. This is due to the lack of access to culture in our settings. This may have led to a lack of PTB confirmation in some of the patients included. Although this is a shortcoming in research, it does, however, reflect real-life conditions and even when culture is included, up to 25% of patients are not microbiologically confirmed.

As apparent from the CXRs shown in the appendix, the films were marked with text and numbers. qXR includes a preprocessing module that zooms in on the relevant portion of the CXR. This removes postprocessing artifacts outside of the lung parenchyma, meaning that writing on the CXRs did not affect the analysis.

The CXR films were photographed using different digital cameras in different settings with varying resolutions. This may introduce potential inconsistencies in image quality and influence analysis. Although this is not optimal in a study setting, it does, however, reflect reality. Two photographs were rejected by both the radiologists and the CAD CXR due to the low quality of the original physical CXR film, whereas the difference in resolution was of no concern for neither radiologist nor CAD CXR. The share of pictures of CXRs from TB-diagnosed patients did not differ between the cameras and the sample from each camera is insufficient to analyze for variability different cameras introduce. We are, however, aware of the potential bias and are currently carrying out a study to investigate this further.

Although more recent recommendations suggest using more costly diagnostic methods such as Xpert up front,¹¹ this strategy may still miss an important group of patients for whom CXR may be a good diagnostic option such as those unable to produce adequate sputum. Moreover, AI-guided CXR interpretation via mobile phones may guide health care workers with no or limited training in assessing CXRs for possible TB.

In conclusion, this pilot study showed that CAD CXR has diagnostic value for PTB comparable with that of CXR interpreted by trained radiologists, even when applied to analog CXRs from patients with TB indicative symptoms photographed by mobile phones or digital cameras. Thus, the use of CAD of PTB has the potential to lessen the currently high case detection gap, particularly in low-resource settings where radiologists are scarce and digital CXRs are often unavailable.

Relationship Disclosure

Qure.AI has not taken part in planning, design, or analysis in connection with the present study. The analysis of the CXRs by the AI (qXR) was done free of charge. Robert, Sridhar, and Tadepalli are employees of Qure.ai. Dr Wejse reports institutional grants from Novo Nordisk Foundation (5 mio kr for “CohereMig - Coherence and co-creation in patient care coordination for migrant patients”; NNF23OC0082779) and Innovation Foundation (24.5 mio kr for “Saliva-based sOlution foR TB diagnosis [SORTS]”; IFD 3146-00019B); reports patent application ongoing for the invention (PCT/EP2023/080683; Diagnosis of tuberculosis in saliva [TECH-2022-631-459]); is the Chairperson Danish Society for Migrant Health, 2023-2025, and Chair of ESGITM (ESCMID Study Group for Infections in Migrants and Travelers) 2022 to ongoing; and stock or stock options in NovoNordisk. The other authors report no conflicts of interest.

Ethics Statement

Approval was obtained from the Guinean Ministry of Health and the Ethics review board of the University of Gondar. Consultative approval was obtained from the Central Ethical Committee in Denmark. All patients provided written informed consent to participate in the study and were offered HIV testing with pretest and posttest counseling. HIV-infected individuals were accompanied to nearby Anti-Retroviral Treatment centers.