Implementation of Digital Chest X-Ray with Computer-Aided Detection for Tuberculosis Screening Among Persons with Advanced HIV Disease in Maputo, Mozambique: A Retrospective Cohort Analysis

1. Introduction

1.1. The Dual Epidemic of TB and HIV in Mozambique

Mozambique faces a severe dual epidemic of HIV and tuberculosis (TB). With an estimated population of 33 million, the country has an adult HIV prevalence of 12.5% and an estimated TB incidence of 361 per 100,000 population [1]. HIV dramatically increases the risk of developing active TB, accelerates its progression, and complicates its diagnosis and treatment. Among people living with HIV (PLHIV), those with advanced HIV disease (AHD) defined by the World Health Organization (WHO) as a CD4 cell count <200 cells/mm³ or the presence of a WHO stage 3 or 4 defining illness are at the highest risk of mortality, with TB being the leading cause of death [2,3].

Despite scale-up of antiretroviral therapy (ART), late presentation to care for HIV and recurrent disengagement from treatment continue to pose significant challenges. In Mozambique, early mortality after ART initiation is high, often driven by undiagnosed opportunistic infections (OIs), principally TB [4]. The national TB case notification rate (334/100,000) closely mirrors the estimated incidence, suggesting a strong detection effort, yet gaps persist, particularly among high-risk subgroups like patients with AHD [1]. The low proportion of bacteriologically confirmed TB (43%) indicates substantial reliance on clinical diagnosis, which can be non-specific in AHD, leading to both over- and under-treatment [1].

1.2. Challenges in TB Diagnosis Among Persons with AHD

Diagnosing TB in AHD is notoriously difficult. Classic symptoms may be absent, and the radiological presentation on chest X-ray (CXR) can be atypical, non-cavitary, and disseminated, often resembling other pulmonary conditions common in HIV— or the chest X-ray may appear entirely normal [3,5]. Sputum-based diagnostic tests, such as Xpert MTB/RIF, have lower sensitivity in this population, especially in those with low CD4 counts and extrapulmonary or paucibacillary disease [6,7]. This diagnostic complexity necessitates a high index of suspicion and a comprehensive screening approach that integrates clinical, radiological, and microbiological tools.

1.3. The Role of Digital Chest X-Ray and Computer-Aided Detection

Digital chest X-ray (dCXR) has emerged as a critical screening tool for TB. It is fast, non-invasive, and can detect abnormalities suggestive of pulmonary TB even in subclinical cases. When paired with artificial intelligence (AI) in the form of computer-aided detection (CAD) software, dCXR can automate image interpretation, providing a standardized, reproducible “score” or probability of TB. This is particularly valuable in settings with a shortage of skilled radiologists, by helping to scale and decentralize screening and prioritize individuals for further diagnostic evaluation [8].

The ideal CAD score threshold may vary by setting, depending on clinical population characteristics, the background burden of non-TB lung disease, and the intended role of CAD in the diagnostic pathway. The WHO now recommends CAD as an alternative to human reading for screening and triage [9].

However, its performance and optimal implementation in real-world, high-HIV-burden clinical settings outside of controlled study conditions require further evaluation. Key questions remain about the optimal CAD score threshold for a given setting, its performance in AHD populations where TB presentation and radiological findings are frequently atypical — or entirely absent —and the practical challenges of integrating this technology into routine clinical workflows.

1.4. The CRAM Clinic as a Natural Implementation Setting

The Centro de Referência do Alto Maé (CRAM), managed by the International Training and Education Center for Health (I-TECH) in Maputo, serves as a centre of excellence for the management of AHD in Mozambique. It provides rapid, comprehensive, same-day screening and initiation of treatment for major OIs, including Cryptococcal meningitis [10] and TB [11,12]. The clinic’s patient population consists of individuals referred with very low CD4 counts, ART treatment failure requiring HIV-1 genotyping [13], or severe comorbidities, and represents a group at extreme risk for TB and mortality [14,15,16]. In 2023, with USAID support, CRAM installed a dCXR machine and CAD software, making it the only primary-level outpatient facility in Maputo city with this capacity.

1.5. Study Rationale and Objectives

Recent studies have evaluated CAD among PLHIV in various settings. Burke et al. demonstrated that CAD combined with urine LAM improved inpatient TB diagnosis [17]. MacPherson et al. found CAD screening increased TB case detection among adults with cough in Malawi [18]. Olotu et al. reported high CAD uptake in Mozambican prisons [19]. Moodley et al. and Kagujje et al. evaluated CAD performance in primary care and among persons with prior TB, respectively [20,21]. Qin et al. and Nzimande et al. provided multi-country and regional validation data [22,23].

However, none of these studies focused specifically on persons with advanced HIV disease (CD4 <200 cells/mm³ or WHO stage 3/4), a group with atypical radiological presentations and particularly high TB burden. Moreover, while efficacy studies have demonstrated CAD’s accuracy in controlled settings, data on its real-world effectiveness within routine, resource-constrained TB/HIV programs remain scarce. Our study addresses both gaps by providing the first dedicated evaluation of CAD performance and threshold selection in an AHD cohort in a high-burden African setting, using routinely collected programmatic data. Guided by implementation science frameworks that emphasize understanding context, processes, and outcomes [24], this retrospective evaluation:

• Described the integration process of dCXR/CAD into the routine clinical workflow of a specialized AHD clinic.
• Evaluated the diagnostic yield of dCXR/CAD among patients with AHD, stratified by CAD result category.
• Assessed the performance (sensitivity and specificity) of the nationally adopted CAD threshold (≥0.5) in the AHD population.
• Listed operational challenges, facilitators, and insights from the first year of implementation.

This analysis provides critical programmatic evidence to inform national policy on CAD use, optimize clinical algorithms for TB screening in AHD, and improve implementation strategies for novel diagnostic technologies in low-resource settings.

2. Materials and Methods

2.1. Evaluation Design and Setting

This was a retrospective cohort analysis of routinely collected clinical and programmatic data. The study was conducted at the CRAM clinic, a specialized outpatient service for AHD within the Alto Maé Health Center in Maputo City, Mozambique. CRAM operates as a referral center for complex HIV cases from across the city’s primary health facilities and hospitals [11].

2.2. Evaluation Population and Cohort

The analysis focused on one cohort screened for TB over a 12-month period (October 2023 – September 2024), consisting of all new HIV+ patients admitted to CRAM during the study period (N=487). Admission criteria included CD4 count <100 cells/mm³ (or <200 if symptomatic), Kaposi’s sarcoma (KS), second-line ART failure, or complex comorbidities. To contextualize CAD performance, we used data from 2,695 patients undergoing dCXR at Alto Maé Health Centre — a primary-level facility serving both HIV-positive and HIV-negative patients — as a proxy for the general population undergoing CXR in this setting. This reference cohort includes the AHD patients described in this study, as both groups used the same machine during the same period.

2.3. The Intervention: dCXR with CAD Integration

The intervention was the integration of systematic dCXR screening with CAD interpretation into the clinic’s existing AHD care package.

A standalone digital X-ray machine was connected to qXR v4.0 (Qure.ai, India), one of six CAD software systems approved by the WHO for TB screening [25]. The CAD software analyses posterior-anterior dCXR images and assigns a score from 0 to 1, indicating the probability of TB-associated abnormalities. According to WHO TDR guidance on CAD calibration studies [25], this represents a triage threshold intended to prioritize individuals for confirmatory testing. The ≥0.5 threshold is the manufacturer’s default setting (qXR v4.0). The National TB Program (PNCT) adopted this threshold pending local calibration. We evaluated the performance of this threshold against recorded diagnosis of TB.

For AHD cohort, dCXR was part of the routine same-day admission assessment. The CAD result was available to clinicians alongside other diagnostic inputs (clinical assessment, Xpert MTB/RIF on sputum or other samples, and urine TB-LAM (performed according to national guidelines [recommended for all PLWH with CD4 <200 cells/mm³ in the outpatient setting). The final decision to initiate TB treatment was made by the clinical team based on a composite assessment.

2.4. Data Sources and variables

Data were extracted from the clinic’s electronic medical records and paper-based registers, which are part of routine national HIV and TB monitoring systems.

Independent variables: CAD result category (Invalid, Normal, Abnormal not suggestive of TB [score <0.5], Abnormal suggestive of TB [score ≥0.5]).

Primary outcome: Recorded diagnosis of TB (bacteriologically confirmed or clinically diagnosed).

Secondary outcome: bacteriological confirmation of TB (by Xpert MTB/RIF or culture).

2.5. Data Analysis

Descriptive statistics were used to characterize the cohort. Analysis was primarily focused on diagnostic yield and CAD performance metrics. For each CAD result category, we calculated: (1) Proportion of patients with recorded diagnosis of TB; and (2) Proportion of those with bacteriological confirmation. We also calculated sensitivity, specificity, PPV, and NPV for the CAD ≥0.5 threshold using the recorded diagnosis of TB as the reference standard.

Traditional sensitivity/specificity calculations against a perfect gold standard were not feasible, because most patients had a clinical diagnosis of TB and did not undergo exhaustive confirmatory testing (e.g., culture). Instead, the proportion of patients ultimately diagnosed and treated for TB was used as a pragmatic measure of CAD’s programmatic utility.

Data were managed and analysed using Microsoft Excel and Stata version 18.0.

2.6. Ethical Considerations

Ethical clearance was obtained from the Mozambican National Bioethics Committee for Health (IRB0002657 - Comité Nacional de Bioética para a Saúde, nr: 21/CNBS/2025) and the permission to perform this evaluation was also obtained from the Health Service of Maputo city (Serviço de Saúde da Cidade, N/Ref. no. 7002/2103/SSCM/2024).

3. Results

3.1. Context and Cohort Description

Between October 2023 and September 2024, 487 new patients were admitted to CRAM. The cohort comprised 53% men, with a median CD4 count at admission of 211 cells/mm³.The main reasons for admission were low CD4 count (40%), cryptococcal disease (11%), Kaposi’s sarcoma (23%), treatment failure (10%), and other medical complications (16%). The median CD4 of 211 cells/mm³ is higher than typically expected in an AHD cohort, which is partly explained by the fact that 23% of patients were admitted for Kaposi’s sarcoma — condition that defines AHD by WHO clinical stage regardless of CD4 count. Notably, 150 out of 487 (32%) of all new enrolees were ultimately diagnosed with TB and started on treatment, underscoring the high TB burden in this population. (Figure 1).

3.2. dCXR/CAD Screening Uptake

Among 487 new AHD patients, 238 (49%) underwent dCXR with CAD reading. The remaining 51% either did not receive a dCXR or their dCXR was not processed by the CAD software due to operational constraints (Figure 1).

3.3. dCXR/CAD Performance and TB Detection

3.3.1. dCXR/CAD Findings in the General Population

In the general population (both PLHIV and HIV-negative patients) screened at the nearby primary health facility (N=2,695), 90% (2,425/2,695) had valid dCXR/CAD results. Of those with a valid result, 43% (1,035/2,425) were interpreted as abnormal. Among these abnormal results, 54% (564/1,035) were flagged as suggestive of TB (CAD score ≥0.5), representing 23% (564/2,425) of all valid dCXR/CAD. This cohort provides a reference for CAD performance in the general population accessing radiography services in Maputo City. As both groups used the same machine during the same period, the AHD patients described in this study are included within this reference cohort; the comparison therefore serves only as a local reference for CAD positivity rates in the Maputo context, rather than a formal comparison between the two cohorts.

3.3.2. TB Detection by CAD Result Category in AHD Patients at CRAM

Among the 238 AHD patients screened with CAD at CRAM: 60% (143/238) had a normal CXR (score typically <0.1); 11% (27/238) had an abnormal CXR not suggestive of TB (score <0.5); 29% (68/238) had a CXR suggestive of TB (score ≥0.5).

TB diagnosis rates increased with CAD score (Table 2): 30% (43/143) among those with a normal result, 52% (15/27) among those with an abnormal non-suggestive result, and 85% (58/68) among those with a suggestive result.

Bacteriological confirmation was similarly low across all groups: 19% (8/43) in the normal group, 20% (3/15) in the non-suggestive group, and 22% (13/58) in the suggestive group. This confirms the high rate of clinical diagnosis of TB in AHD population (Table 1).

3.4. Diagnostic Performance of CAD ≥0.5 Threshold

The diagnostic performance of the manufacturer-recommended CAD threshold (≥0.5) was evaluated against the recorded diagnosis of TB. Table 2 presents the 2×2 contingency table with case and control classifications.

Table 2. 2×2 contingency table – CAD ≥0.5 threshold vs. TB diagnosis (reference standard).

	TB Case (Recorded diagnosis of TB)	Control (Not diagnosed with TB)	Subtotal
CAD Suggestive (≥0.5)	58 (True Positive)	10 (False Positive)	68
CAD Not Suggestive (<0.5)	58 (False Negative)	112 (True Negative)	170
Subtotal	116	122	238 (Total)
Threshold ‘suggestive of TB’ = CAD score ≥0.5

Table 2. 2×2 contingency table – CAD ≥0.5 threshold vs. TB diagnosis (reference standard).

	TB Case (Recorded diagnosis of TB)	Control (Not diagnosed with TB)	Subtotal
CAD Suggestive (≥0.5)	58 (True Positive)	10 (False Positive)	68
CAD Not Suggestive (<0.5)	58 (False Negative)	112 (True Negative)	170
Subtotal	116	122	238 (Total)
Threshold ‘suggestive of TB’ = CAD score ≥0.5

From Table 2, the following performance metrics were calculated: Sensitivity;

Specificity; Positive Predictive Value (PPV); and Negative Predictive Value (NPV)

(Table 3)

3.4.1. Interpretation of Table 2 and Table 3:

At the current threshold of ≥0.5, CAD missed 58 of 116 TB cases (false negatives), resulting in a sensitivity of only 50%. However, among the 68 patients flagged as CAD-positive, 58 (85%) were true positives, highlighting a relatively high PPV (85%). The relatively low NPV (66%) indicates that a negative CAD result does not reliably rule out TB in this AHD population. The high specificity (92%) means that false positives are relatively less common (8%).

The mean CAD score across all patients varied by result category: 0.07 among those with a normal result (n=143), 0.34 among those with an abnormal but non-suggestive result (n=27), and 0.83 among those with a result suggestive of TB (n=68).

4. Discussion

This retrospective analysis of the first year of dCXR/CAD implementation at a reference AHD clinic in Mozambique provides real-world evidence on the promise and pitfalls of this technology in a high-burden, resource-constrained setting. The discussion is framed within the RE-AIM core domains of implementation science (Reach, Effectiveness, Adoption, Implementation, Maintenance) [24]

4.1. Reach and Adoption: Integrating Technology Into Routine Care

The reach of CAD screening among AHD patients was 49%, indicating that while the tool was adopted, its consistent application was limited. For AHD patients, the integration was into a complex, multi-faceted same-day assessment for high risk and sometimes critically ill individuals. Operational barriers such as the non-integrated system and lack of dedicated staff directly limited reach. This underscores a key implementation principle: even highly effective tools fail if the workflow integration is suboptimal [17]. Successful adoption requires not just technology but also process redesign and dedicated human resources to manage it.

4.2. Effectiveness: CAD Performance in Context: Confirmatory Rather than Triage Tool

The performance of CAD in this cohort reflects a pattern more consistent with a confirmatory adjunct than a triage tool. In a population where advanced HIV disease and a high background TB burden converge, the pre-test probability of TB is naturally elevated. A positive CAD result — with a PPV of 85%— adds meaningful diagnostic weight when clinical evaluation already points towards TB, supporting the decision to initiate treatment. However, the low sensitivity of 50% means that a negative result carries limited reassurance. Among AHD patients, 30% of those with a normal CAD result and 52% of those with an abnormal but non-suggestive result were ultimately diagnosed with TB. This reflects a well-recognized phenomenon: TB presentation in severely immunocompromised patients is frequently atypical — subtle, disseminated, or mimicking other opportunistic infections — and CAD algorithms trained on general population data may lack sensitivity for these patterns [5,10]. A non-suggestive CAD result should therefore not reduce clinical suspicion in a patient with advanced HIV disease.

4.3. Clinical Diagnosis Requires a Multi-Test Approach

Importantly, diagnosis in this population must rest on a constellation of findings rather than any single result. The low bacteriological confirmation rates in this cohort (19–22%) reinforce that clinicians need to integrate clinical signs, symptoms, and complementary tests — including urine TB-LAM — alongside CAD output [26].

4.4. Threshold Recalibration and Risk-Stratified Implementation in AHD

Among patients with an abnormal but non-suggestive result, the mean CAD score was 0.34 and 52% were ultimately diagnosed with TB, providing empirical grounds for reconsidering the current threshold in this population. Whether a lower threshold could reposition CAD closer to a triage role remains an important question. Lowering the threshold from 0.5 to 0.4, for example, would be expected to improve sensitivity — capturing a greater proportion of true TB cases — at the cost of reduced specificity and a likely decline in PPV. This trade-off could bring CAD’s performance profile more in line with the requirements of a triage tool, where the priority is to minimize missed cases and identify who needs confirmatory investigation, accepting a higher rate of false positives. Our data do not allow us to answer this question, but threshold optimization in this specific population deserves further investigation. Context-specific calibration of AI thresholds remains a recognised challenge in global health implementation [27]

The findings highlight the tension between fidelity to the manufacturer-recommended threshold (≥0.5) and the need for local adaptation [28]. For AHD patients, a lower CAD threshold combined with a predefined intensive workup protocol — Xpert, urine TB-LAM, ultrasound — could formalize this adaptation [29].

4.5. Challenges and Sustainability (Maintenance)

The identified challenges point to several threats to long-term sustainability. First, system fragmentation represents a critical bottleneck, as the disconnect between the X-ray machine and the CAD software necessitates manual image transfers and creates dependence on stable connectivity; therefore, investment in integrated, plug-and-play systems is important for scaling [30]. Second, a human resource gap exists because technology does not replace human input but rather reallocates it. A designated staff role for dCXR and CAD management is critical for consistent use and data quality [9]. Third, the lack of quality assurance systems poses a significant risk. For CAD to be trusted and improved, a national monitoring system and periodic recalibration against local patient outcomes are needed [22,25].

4.6. Limitations

This study has limitations inherent to retrospective programmatic data: missing data (especially for patients not receiving CAD), potential variability in clinical diagnostic criteria between clinicians, and the absence of a culture-confirmed gold standard. The sensitivity and specificity estimates we report are therefore reference-dependent and may overestimate performance if clinical diagnoses include false positives, or underestimate it if subclinical cases were missed. Patients referred for dCXR/CAD appeared to have a higher pretest probability for TB (49% TB diagnosis vs. 14% among those not screened with CAD), suggesting a selection bias. Importantly, this bias likely inflates PPV estimates, meaning that CAD’s confirmatory value may be lower in less selected populations. Whether the low sensitivity we report would improve or worsen under more systematic and less selective application of CAD remains uncertain and can only be answered through prospective evaluation with broader screening coverage. Finally, our findings apply to adults with AHD in a single urban site; generalizability to children with HIV or to more remote, lower-resource settings requires separate evaluation, though CRAM is likely representative of urban AHD cohorts in high-burden African settings [22,25].

5. Conclusions and Recommendations

The implementation of dCXR/CAD at CRAM confirms the high yield of systematic TB screening in AHD populations, while also revealing how its current application — using a general-population threshold — risks missing a significant number of TB cases among the most vulnerable patients while generating false positives in others. Based on these results and our experience implementing this technology, we recommend:

• For national programs: Conduct an urgent review of the CAD score threshold for PLHIV, particularly those with CD4 <200 cells/mm³. Consider establishing a lower, more sensitive threshold for this group.
• For clinic managers: Integrate dCXR/CAD findings into a composite clinical algorithm for AHD. A non-suggestive CAD result should not end the screening process for AHD patients. Advocate for and allocate a dedicated staff position to manage the CAD workflow.
• For implementers and donors: Prioritize funding for integrated dCXR/CAD systems over fragmented components. Support the development of national QA systems and operational research to continually evaluate and adapt CAD use in different patient populations.
• For future research: Prospective studies are needed to validate an optimized, HIV-specific CAD threshold and to evaluate the cost-effectiveness and impact on mortality of dCXR/CAD-based screening in AHD care packages.

In conclusion, dCXR/CAD shows real promise as a decision-support tool in AHD care, but its current performance in this population falls short of what is needed for either triage or confirmation. Realising its potential will require threshold optimization, careful clinical integration, and a commitment to evaluate global tools against local, high-risk realities.