Diagnostic accuracy of an artificial intelligence-based osteoporosis screening system on portable chest radiographs Authors: Jee Hyun Kim, So Hyun Ahn, Rena Lee, Sungho Cho, Soeun Choi, Kwan-Chang Kim Published: 2026 ABSTRACT BackgroundArtificial intelligence (AI) models applied to conventional chest radiographs (CXRs) have shown potential for osteoporosis screening in hospital settings. However, their performance when applied to portable CXR obtained in community programs remains uncertain. This study evaluated a commercially available AI model (PROS® CXR: OSTEO) for osteoporosis screening using portable CXR.MethodsOlder adults participating in a Korean National Tuberculosis Association mobile screening program who consented to osteoporosis evaluation were prospectively enrolled. Portable CXR images were analyzed by the AI model to generate a continuous osteoporosis risk score (0–1). Participants subsequently underwent dual-energy X-ray absorptiometry (DXA) within approximately one week. Osteoporosis was defined as a T-score ≤−2.5 at the lumbar spine, femoral neck, or total hip. Diagnostic performance was assessed across thresholds from 0.0 to 1.0 to determine an optimal threshold.ResultsFifty-two participants with paired portable CXR and DXA data were analyzed. Osteoporosis was present in 20 participants (38.5%). A threshold of 0.2 provided the most favorable screening performance in this cohort, yielding an accuracy of 0.69, sensitivity of 0.90, specificity of 0.57, positive predictive value of 0.56, negative predictive value of 0.90, and an F1-score of 0.69. The area under the curve was 0.86 (95% confidence interval: 0.73–0.97). Most false-positive cases occurred in osteopenic individuals, whereas false negatives were limited to borderline osteoporosis cases.ConclusionsIn this feasibility-based pilot study, the AI model demonstrated promising discrimination when applied to portable CXR in community-based osteoporosis screening. Its high sensitivity at the selected threshold suggests potential utility as a referral-oriented screening tool, but the findings require confirmation in larger, more balanced, externally validated cohorts. Highlight box Key findings • In this feasibility-based pilot study, a commercially available artificial intelligence (AI)-based osteoporosis screening system showed preliminary discriminatory performance when applied to portable chest radiographs (CXRs) acquired during community-based mobile tuberculosis screening. What is known and what is new? • AI-based osteoporosis screening using conventional fixed CXRs has shown promising diagnostic performance in previous studies. • This study evaluated the same type of AI-based screening approach in portable CXRs obtained in a community mobile screening setting, where image acquisition conditions may differ from standard hospital-based radiography. What is the implication, and what should change now? • Portable CXR-based AI may have potential as a referral-oriented screening tool to identify older adults who may benefit from confirmatory dual-energy X-ray absorptiometry testing. • However, because of the small and sex-imbalanced cohort, lack of systematic image-quality assessment, limited clinical characterization, and absence of external validation, the findings should be considered preliminary and hypothesis-generating. Larger, clinically characterized, multicenter validation studies are needed before routine implementation. Introduction Osteoporosis is one of the most common chronic conditions associated with aging, characterized by reduced bone mass and deterioration of bone microarchitecture, which significantly increase the risk of fractures. Vertebral and hip fractures, in particular, are not only associated with pain and disability but also with long-term functional decline, reduced quality of life, institutionalization, and excess mortality (1,2). According to the World Health Organization (WHO) and international societies, the socioeconomic burden of osteoporotic fractures continues to rise, and by 2050, more than half of all hip fractures are projected to occur in Asia (3,4). Dual-energy X-ray absorptiometry (DXA) remains the gold standard for osteoporosis diagnosis, providing quantitative assessment of bone mineral density (BMD). However, the high cost of equipment, need for specialized infrastructure, and reliance on trained personnel limit its availability, especially in primary care settings, rural communities, and resource-limited environments (5,6). Consequently, many patients remain undiagnosed and untreated, leading to preventable morbidity and mortality. To overcome these diagnostic gaps, artificial intelligence (AI) has been increasingly applied to medical imaging. Chest radiographs (CXRs), one of the most widely performed imaging modalities worldwide, represent a promising tool for opportunistic screening of osteoporosis. Because CXRs are routinely obtained for pulmonary and cardiovascular evaluation, AI-driven analysis can provide additional insights into skeletal health without additional radiation exposure or cost. Several recent studies have demonstrated the feasibility of this approach. For instance, Sato et al. [2022] developed a deep learning model predicting T-scores from CXRs, achieving an area under the curve (AUC) of 0.84 for the diagnosis of osteoporosis (T-score ≤−2.5) (1). Wang et al. [2021] proposed a multi-region-of-interest (multi-ROI) CXR-based model and reported excellent discrimination performance with an AUC of 0.936 (3). Similarly, Pan et al. [2020] utilized low-dose chest computed tomography (LDCT) scans for opportunistic BMD estimation, reporting AUC values of 0.927 for osteoporosis detection and 0.942 for low bone mass detection (5). Comparable computed tomography-based opportunistic screening approaches have likewise reported high diagnostic performance (7-9). Despite these advances, most studies have relied on fixed hospital-based imaging equipment, limiting their applicability in real-world scenarios where access to DXA or advanced imaging is scarce. For older adults, frail individuals, or residents in rural and long-term care facilities, accessibility to such diagnostic tools remains a major challenge. In this context, portable CXRs systems offer a unique opportunity to bridge this gap. Portable X-ray units are already widely used in public health campaigns, community health centers, and mobile screening programs such as tuberculosis detection. They allow immediate, bedside imaging without requiring patients to travel to tertiary care facilities, making them an attractive platform for opportunistic osteoporosis screening (10). The integration of AI-based diagnostic algorithms with portable CXR has two major implications. First, it enhances accessibility by enabling opportunistic screening even in underserved settings, thereby improving equity in osteoporosis care. Second, it increases efficiency, as a single portable CXR examination could simultaneously serve dual purposes: screening for pulmonary disease (e.g., tuberculosis) and osteoporosis. This multipurpose approach maximizes healthcare resource utilization and may provide a cost-effective strategy for early intervention (11,12). In this study, we evaluated the diagnostic accuracy of an AI-based osteoporosis detection program applied to portable CXR with DXA serving as the reference standard. We aimed to evaluate whether portable CXR combined with AI could provide preliminary evidence for opportunistic osteoporosis screening in community-based mobile screening settings. We present this article in accordance with the STARD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0668/rc). Methods Study design This was a single-site, prospective diagnostic-accuracy, feasibility-based pilot study conducted as a commissioned project to develop an AI-assisted osteoporosis screening platform using a portable X-ray system. Because the sample size was determined by feasibility within the study period, the findings were intended to provide preliminary and hypothesis-generating evidence to inform future validation studies. The present study was designed to evaluate the AI system as an image-only screening tool based on portable CXRs. The study was carried out in collaboration with the Ulsan-Gyeongnam branch of the Korean National Tuberculosis Association (KNTA). Data were collected at Jindong Community Welfare Center in Masanhappo-gu, Changwon, Gyeongsangnam-do, Korea, where community-dwelling older adults underwent portable CXR as part of a public health screening service. Within this setting, eligible residents who agreed to participate received AI-based osteoporosis screening on their portable CXR, followed by referral for DXA. The present analysis evaluated the diagnostic performance of the AI-based portable CXR index test using DXA as the reference standard. Participants and data collection All participants in this study were community-dwelling residents registered for a vulnerable-population health screening at Jindong Community Welfare Center. During the screening sessions run by the Ulsan-Gyeongnam branch of the KNTA between September and October 2024, residents aged 60 years or older were informed about an additional osteoporosis substudy. Those who wished to participate and provided written informed consent were enrolled. A total of 52 individuals met these conditions and constituted the study cohort. As part of the routine screening program, all enrolled participants had already undergone portable CXR for tuberculosis screening, and these existing CXR images were later used as input to the AI software to generate an osteoporosis risk score (index test). After enrollment, participants were referred to Cheonga Hospital for a DXA examination, which was scheduled to take place within approximately one week of the CXR visit and served as the reference standard. All 52 participants completed both the AI-based index test and the DXA examination and were included in the final analysis. For each participant, demographic and anthropometric variables (sex, year of birth, age, height, and weight), as well as BMD and T-scores at the lumbar spine, femoral neck, and total hip, were obtained from the DXA report issued by Cheonga Hospital. The AI score (continuous value between 0 and 1) corresponding to each portable CXR was exported from the ProMedius PROS® CXR: OSTEO platform (ProMedius, Seoul, Korea). All variables were compiled into an anonymized analysis dataset for statistical analysis. Image and DXA acquisition All CXRs used in this study were pre-existing images obtained with a portable X-ray system as part of the routine tuberculosis screening service conducted at a single community screening center, Jindong Community Welfare Center, by the Ulsan-Gyeongnam branch of the KNTA. To minimize technical variability, all images were acquired within the same screening program using the same portable X-ray unit, detector, and standardized acquisition protocol. Participants did not undergo any additional chest imaging specifically for this study; the same portable CXRs that had been acquired for tuberculosis screening were secondarily used as input for the AI-based osteoporosis assessment. Portable radiography was performed using a REMEX-KA6 X-ray unit with a tube voltage of 70 kVp and a maximum tube current of 6 mA (Figure 1). Images were acquired with a flat-panel detector (VIVIX-S 1717V, Vieworks) with a pixel pitch of 140 µm and an active area of 43 cm × 43 cm. All examinations were obtained in a posteroanterior projection at a source-to-detector distance of approximately 1.5 m (Table 1). Portable chest radiograph unit and detector. Photograph of the portable X-ray system (REMEX-KA6) and flat-panel detector (VIVIX-S 1717V, Vieworks) used for community-based chest radiography in this study. The system was used to acquire posteroanterior chest radiographs during mobile tuberculosis screening. Table 1Portable chest radiograph acquisition parametersParametersValuesTube voltage (kVp)70Tube current (mA)6Exposure time (s)0.65Source-to-image distance (m)1.50Detector pixel size (µm)140 Portable chest radiograph acquisition parameters All included portable CXRs were considered diagnostically acceptable for tuberculosis screening according to the KNTA workflow. However, no structured or quantitative image-quality assessment was performed specifically for AI-based osteoporosis analysis. Patient positioning, exposure variability, motion, rotation, and radiographic artifacts were not systematically scored in the present study. DXA was performed at Cheonga Hospital using a GE Lunar DPX densitometer (GE Healthcare, Chicago, IL, USA) according to the manufacturer’s standard clinical protocol. With the participant in the supine position, BMD was measured at the lumbar spine, femoral neck, and total hip, and T-scores for each region were automatically calculated by the DXA system and reported on the printed DXA report. These DXA-derived T-scores were used to define the reference-standard osteoporosis status, as described in section “AI index test and reference standard labeling”. AI index test and reference standard labeling AI-based index test The index test was a commercially available AI software, ProMedius PROS® CXR: OSTEO, which was developed based on a deep-learning algorithm originally trained at Asan Medical Center to classify DXA-defined osteoporosis from posteroanterior CXRs. In the development study, the convolutional neural network was trained on paired CXRs and central DXA measurements; the model architecture and training procedures have been described in detail elsewhere. In the present study, we used the finalized commercial software as a commercial “black-box” index test without modification, retraining, or recalibration. The internal model architecture, feature attribution, and portable CXR-specific adaptation procedures were not accessible to the investigators. This approach was intended to evaluate the software under real-world deployment conditions, as it would be used in routine community-based screening practice. For each participant, the pre-existing portable CXR acquired for tuberculosis screening was uploaded to the cloud-based PROS® CXR: OSTEO platform in its original DICOM format. The software automatically processed the image and generated a continuous, patient-level osteoporosis risk score between 0 and 1, where higher values indicate a greater likelihood of DXA-defined osteoporosis. The standard clinical inference mode and default image-handling settings recommended by the manufacturer were used; no manual preprocessing, image adjustment, or multiple readings were performed. The resulting continuous AI scores were subsequently compared with DXA-based labels for diagnostic accuracy analyses, as described in sections “Statistical analysis” and “Results”. Reference standard and labeling The reference standard was DXA performed at Cheonga Hospital within approximately one week of the portable CXR examination. DXA was selected as the reference standard because it is the established clinical gold standard for the diagnosis of osteoporosis. It was performed according to routine clinical protocol, and the assessors of the reference standard were not aware of the AI-based index test results. For each participant, BMD and T-scores for the lumbar spine, femoral neck, and total hip were obtained from the printed DXA report. In accordance with the hospital’s DXA reporting protocol and WHO criteria, three composite T-scores were defined. The lumbar T-score was defined as the mean of the T-scores reported for six contiguous lumbar regions: L1–L2, L1–L3, L1–L4, L2–L3, L2–L4, and L3–L4. The femoral neck T-score was defined as the T-score at the femoral neck, and the femoral total T-score was defined as the T-score at the “total” region of the proximal femur. Osteoporosis (reference-positive) was diagnosed when at least one of these three composite T-scores (lumbar, femoral neck, or femoral total) was ≤−2.5. Osteopenia was defined as a T-score between −2.5 and −1.0, and normal bone density as a T-score ≥−1.0. For the primary binary analysis, osteopenia and normal bone density were grouped together as reference-negative. Statistical analysis The primary endpoints were the sensitivity and specificity of the AI index test for detecting DXA-defined osteoporosis at prespecified AI score thresholds (described in section “Results”). Secondary endpoints included the positive predictive value (PPV), negative predictive value (NPV), and overall discrimination summarized by the receiver operating characteristic (ROC) curve. Continuous variables were summarized as mean ± standard deviation (SD) or median with interquartile range (IQR), as appropriate. Categorical variables were summarized as counts and percentages. Diagnostic accuracy measures were derived from 2×2 contingency tables. For each selected AI score threshold, true positives (TPs), true negatives (TNs), false positives (FPs), and false negatives (FNs) were defined using DXA as the reference standard and the thresholded AI output as the index test. Sensitivity, specificity, PPV and NPV were calculated as follows: Each proportion was reported with a two-sided 95% confidence interval (CI) estimated using the Wilson method. AUC was calculated using the continuous AI scores (0–1) against the binary DXA-defined labels without fixing a specific probability threshold, and its 95% CI was estimated using DeLong’s approach. A two-sided 95% CI was used to quantify statistical uncertainty. All analyses were conducted on the complete-case dataset (n=52) using standard statistical software (e.g., R or Python). Ethics This study was conducted in the context of a government-supported public health screening service operated by the Ulsan-Gyeongnam branch of the KNTA at Jindong Community Welfare Center. No additional chest imaging beyond routine tuberculosis screening was performed for research purposes; the pre-existing portable CXR images acquired for tuberculosis screening were secondarily used as input for the AI-based osteoporosis assessment. DXA was offered as an optional, second-stage examination for participants who consented to the osteoporosis substudy. Only de-identified data were used for analysis. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and all analyses were performed using anonymized datasets. In accordance with applicable institutional and national regulations for public health screening projects using anonymized data, formal institutional review board approval was not required. All participants received an explanation of the study and provided written informed consent for the use of their clinical data and images for research purposes. Results Patient characteristics A total of 52 participants were included in the analysis, with a mean age of 77.96±5.14 years (Table 2). The study population was predominantly female, comprising 47 women (90.40%) and 5 men (9.60%). The mean height and weight were 152.67±7.32 cm and 57.12±9.11 kg, respectively. Table 2Baseline characteristics of the study populationCharacteristicsParticipants (n=52)Age (years)77.96±5.14Sex Female47 (90.40) Male5 (9.60)Height (cm)152.67±7.32Weight (kg)57.12±9.11T-scoreLumbar−1.51±1.46Femoral neck−1.47±1.15Total femur−1.33±1.26BMD categories Normal10 (19.20) Osteopenia22 (42.30) Osteoporosis20 (38.50)Data are presented as mean ± standard deviation or n (%). BMD, bone mineral density; SD, standard deviation. Baseline characteristics of the study population Data are presented as mean ± standard deviation or n (%). BMD, bone mineral density; SD, standard deviation. The mean T-scores were −1.51±1.46 at the lumbar spine, −1.47±1.15 at the femoral neck, and −1.33±1.26 at the total femur, indicating generally reduced bone density in this older cohort. Based on DXA, 10 participants (19.2%) had normal bone density, 22 (42.3%) were classified as osteopenia, and 20 (38.5%) met the diagnostic criteria for osteoporosis. Thus, more than three-quarters of the screened population showed either osteopenia or osteoporosis. Threshold optimization for the AI risk score Because the original ProMedius algorithm had been developed and validated on fixed digital radiography systems using a default decision threshold of 0.5, we first applied this operating point to the portable CXR data. At a threshold of 0.5, however, the model did not identify any DXA-defined osteoporosis cases (sensitivity 0.00, 95% CI: 0.00–0.00), while classifying all non-osteoporotic participants as negative (specificity 1.00, 95% CI: 1.00–1.00) and yielding only moderate overall accuracy (0.61, 95% CI: 0.48–0.75) with an F1 score of 0.00 (Table 3, Figure 2). This pattern indicated that the original threshold was overly conservative in the portable screening setting and unsuitable for opportunistic case finding. Table 3Diagnostic performance of the AI model at different score thresholdsThresholdAccuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)F1 score (95% CI)0.00.39 (0.25–0.52)1.00 (1.00–1.00)0.00 (0.00–0.00)0.55 (0.40–0.68)0.10.48 (0.35–0.62)0.95 (0.83–1.00)0.19 (0.06–0.34)0.58 (0.42–0.71)0.20.69 (0.56–0.83)0.90 (0.75–1.00)0.57 (0.38–0.75)0.69 (0.53–0.81)0.30.71 (0.58–0.85)0.35 (0.14–0.56)0.94 (0.85–1.00)0.47 (0.22–0.69)0.40.66 (0.52–0.79)0.10 (0.00–0.25)1.00 (1.00–1.00)0.17 (0.00–0.40)0.50.61 (0.48–0.75)0.00 (0.00–0.00)1.00 (1.00–1.00)0.00 (0.00–0.00)0.60.61 (0.48–0.75)0.00 (0.00–0.00)1.00 (1.00–1.00)0.00 (0.00–0.00)AI, artificial intelligence; CI, confidence interval. Diagnostic performance of the AI model at different score thresholds AI, artificial intelligence; CI, confidence interval. Diagnostic performance across AI risk score thresholds. Accuracy, sensitivity, and specificity are plotted according to threshold values ranging from 0.0 to 1.0. The selected operating threshold (0.2) and the manufacturer’s default threshold (0.5) are highlighted for comparison. AI, artificial intelligence. To explore a more appropriate operating point, we systematically evaluated model performance at AI score thresholds from 0.0 to 1.0 in 0.1 increments. As expected, increasing the threshold led to a monotonic decrease in sensitivity and a corresponding increase in specificity. At very low thresholds (0.0 and 0.1), sensitivity remained extremely high (1.00 and 0.95, respectively), but specificity was poor (0.00 and 0.19), resulting in a large number of false-positive classifications despite only modest accuracy (0.39 and 0.48, respectively). At an AI score threshold of 0.2, the model achieved a more favorable balance between sensitivity and specificity, with an accuracy of 0.69 (95% CI: 0.56–0.83), sensitivity of 0.90 (95% CI: 0.75–1.00), specificity of 0.57 (95% CI: 0.38–0.75), and the highest F1 score among the evaluated cutoffs (0.69, 95% CI: 0.53–0.81). Although the highest absolute accuracy was observed at a threshold of 0.3 (0.71, 95% CI: 0.58–0.85), this came at the cost of a marked reduction in sensitivity to 0.35 (95% CI: 0.14–0.56) despite very high specificity (0.94, 95% CI: 0.85–1.00). Given the screening context and the clinical priority of minimizing missed osteoporosis cases, the 0.2 threshold was considered the most appropriate operating point and was therefore used for the primary analyses presented in this study. Overall diagnostic performance at the selected operating point Figure 3 shows the ROC curve for the AI risk score in discriminating DXA-defined osteoporosis from non-osteoporosis. The model achieved an AUC of 0.86 (95% CI: 0.73–0.97), indicating good overall discriminatory ability across the full range of thresholds. ROC curve of the AI model for DXA-defined osteoporosis. ROC curve illustrating the discrimination performance of the AI risk score for identifying osteoporosis defined by DXA. The AUC with 95% CI is shown. AI, artificial intelligence; AUC, area under the curve; CI, confidence interval; DXA, dual-energy X-ray absorptiometry; ROC, receiver operating characteristic. At the selected operating point with an AI score threshold of 0.2 (highlighted on the ROC curve), the model yielded an accuracy of 0.69 (95% CI: 0.56–0.83), a sensitivity of 0.90 (95% CI: 0.75–1.00), and a specificity of 0.57 (95% CI: 0.38–0.75). The PPV and NPV at this threshold were 0.56 and 0.90, respectively, indicating that AI-positive classifications moderately enriched for true osteoporosis, whereas AI-negative classifications rarely included osteoporosis cases. The corresponding F1 score was 0.69 (95% CI: 0.53–0.81), reflecting a favorable balance between precision and recall at this threshold. Consistent with the threshold analysis in section “Threshold optimization for the AI risk score”, these findings confirm that the 0.2 operating point prioritizes high sensitivity while maintaining only moderate specificity, which is appropriate for community-based opportunistic screening. Analysis of misclassified cases At the selected operating point (AI score threshold 0.2), the confusion matrix comprised 18 TPs, 2 FNs, 18 TNs, and 14 FPs (Figure 4). Review of the misclassified cases helped characterize the model’s error pattern in the portable CXR setting. Confusion matrix at the selected operating threshold (AI score =0.2). Rows indicate the DXA reference standard, and columns indicate the AI prediction. Label 1 denotes DXA-defined osteoporosis or an AI-positive prediction, whereas label 0 denotes non-osteoporosis or an AI-negative prediction. The four cells represent TNs (n=18), FPs (n=14), FNs (n=2), and TPs (n=18). AI, artificial intelligence; DXA, dual-energy X-ray absorptiometry; FNs, false negatives; FPs, false positives; TNs, true negatives; TPs, true positives. Among the 14 false-positive classifications, 12 occurred in participants with osteopenia, defined as a T-score between −2.5 and −1.0 across the lumbar spine, femoral neck, or total femur. Only two false-positive cases were observed in individuals with normal bone density (all three composite T-scores >−1.0). This pattern indicates that the algorithm tended to overestimate osteoporosis risk in patients with reduced but not frankly osteoporotic bone mass, particularly those in the osteopenic range. Clinically, these cases would prompt further DXA evaluation despite not meeting the DXA threshold for osteoporosis. In contrast, two participants with DXA-defined osteoporosis were incorrectly classified as negative by the AI system (FNs). Both had T-scores close to the diagnostic cutoff (borderline osteoporosis), and their AI scores were 0.15 and 0.07, respectively, falling just below the 0.2 threshold. These findings suggest that borderline osteoporosis cases with marginal T-scores (T-scores close to the −2.5 diagnostic cutoff) remain challenging for the AI model, especially when the predicted risk score lies near the decision boundary. Taken together, the misclassification analysis shows that the model favors high sensitivity at the cost of reduced specificity, primarily due to overcalling osteopenia as osteoporosis, while a small number of borderline osteoporosis cases may still be missed at the chosen threshold. Diagnostic performance across BMD categories To further characterize how the AI system behaved across different BMD strata, we evaluated diagnostic performance separately within the normal, osteopenia, and osteoporosis groups at the selected threshold of 0.2 (Table 4, Figure S1). Table 4Exploratory subgroup performance of the AI model across BMD categories at the threshold of 0.2BMD diagnosisAccuracySensitivitySpecificityF1 scoreNormal0.600.000.600.00Osteopenia0.550.000.550.00Osteoporosis0.900.900.000.95AI, artificial intelligence; BMD, bone mineral density. Exploratory subgroup performance of the AI model across BMD categories at the threshold of 0.2 AI, artificial intelligence; BMD, bone mineral density. In participants with DXA-defined osteoporosis, the AI model showed excellent performance, with an accuracy of 0.90, sensitivity of 0.90, and an F1 score of 0.95. This indicates that the majority of osteoporotic participants were correctly flagged as positive at the chosen operating point. By contrast, performance in the non-osteoporotic groups was more limited. Among individuals with osteopenia, the accuracy and specificity were both 0.55, whereas in those with normal bone density these metrics were 0.60. Sensitivity and F1 score for the normal and osteopenia groups were 0.00 because, in this binary classification framework, no participants in these groups met the definition of the positive class (DXA-defined osteoporosis). Therefore, these measures are of limited value for comparing performance across BMD categories, and accuracy may provide a more interpretable summary for such exploratory subgroup comparisons. These results further suggest that the current model is most useful for identifying DXA-defined osteoporosis rather than discriminating among non-osteoporotic BMD categories. Sex-specific diagnostic performance We next examined whether diagnostic performance differed by sex at the selected AI score threshold of 0.2 (Table 5, Figure S2). In the female subgroup, the model achieved an accuracy of 0.66, with a sensitivity of 0.88 and a specificity of 0.53. The corresponding F1 score was 0.65, indicating reasonably good detection of DXA-defined osteoporosis in women, albeit with a moderate rate of false-positive classifications among non-osteoporotic women. Table 5Diagnostic performance of the AI model by sex at the AI score threshold of 0.2 (male subgroup, n=5)SexAccuracySensitivitySpecificityF1 scoreFemale0.660.880.530.65Male1.001.001.001.00AI, artificial intelligence. Diagnostic performance of the AI model by sex at the AI score threshold of 0.2 (male subgroup, n=5) AI, artificial intelligence. In contrast, the performance metrics in the male subgroup were uniformly 1.00 for accuracy, sensitivity, specificity, and F1 score. Because only five male participants were included, the apparently perfect performance observed in men should be regarded as descriptive only and not interpreted as evidence of robust sex-specific performance. Because of the very small number of male participants, sex-specific performance estimates should be regarded as exploratory and descriptive only. Larger studies with more balanced sex distributions are needed to determine whether any apparent sex-related differences in model performance are reproducible and clinically meaningful. Discussion In this community-based pilot study, we evaluated a deep learning-based osteoporosis screening algorithm (PROS® CXR: OSTEO) applied to portable CXR acquired during a mobile tuberculosis screening program. Using DXA as the reference standard, the AI model achieved good overall discrimination (AUC 0.86) with high sensitivity (0.90) and moderate specificity (0.57) at an optimized threshold of 0.2. These findings suggest that, even in the portable imaging setting, the model retained a degree of discrimination that may support referral-oriented osteoporosis screening. Our results extend prior work on the development and validation of chest radiograph-based osteoporosis models using fixed digital radiography systems. Jang et al. developed the OsPor-screen model using more than 13,000 paired DXA and CXRs and reported AUCs of 0.91 and 0.88 in internal and external test sets, respectively, for DXA-defined osteoporosis (lumbar spine, femoral neck, or total hip T-score ≤−2.5) (13). In a subsequent clinical cohort study using the commercial PROS® CXR: OSTEO software, Koo et al. observed excellent performance, with AUCs of 0.94 for the femur and 0.93 for the lumbar spine and sensitivity and specificity of 90% and 81%, respectively, for osteoporosis screening (14). Compared with these fixed-hospital radiography studies, our portable CXR-based AUC of 0.86 is modestly lower and specificity is clearly reduced, but the sensitivity of 0.90 is in a similar range. This suggests that much of the diagnostic signal learned from high-quality fixed CXRs can still be leveraged in lower-dose and more variable portable images, albeit with a trade-off in false-positive rates. The threshold analysis in our cohort provides additional insight into how PROS® CXR: OSTEO should be operated in resource-limited, screening-focused environments. In the JTD clinical cohort, PROS® CXR: OSTEO was evaluated using a predefined decision boundary of 0.2, which yielded a balanced combination of high sensitivity and moderate specificity in conventional posteroanterior radiographs (14). In our portable setting, however, the manufacturer’s default threshold of 0.5 essentially eliminated sensitivity, whereas systematic exploration of thresholds from 0.0 to 1.0 demonstrated that 0.2 offered the best compromise between avoiding missed osteoporosis and containing the number of unnecessary referrals. At this threshold, the model achieved the most screening-appropriate balance between sensitivity and specificity while also yielding the highest F1 score among the evaluated cutoffs. Our data suggest that the original operating threshold may not be directly transferable to the portable screening setting and that the operating threshold may need to be re-evaluated when the acquisition environment and target population differ from those of the original development datasets. Misclassification analysis highlighted patterns that are consistent with previous CXR- and CT-based opportunistic screening studies. In our cohort, most false-positive cases occurred in participants with osteopenia rather than normal bone density, indicating that the AI model tends to “overcall” borderline low bone mass as osteoporosis. Jang et al. and other opportunistic imaging studies likewise reported enrichment of osteopenia among false-positive classifications, reflecting the intrinsic difficulty of sharply separating osteopenia from osteoporosis when T-scores lie near diagnostic cut-offs (13,15). Conversely, the two false-negative cases in our study both had borderline osteoporosis with T-scores close to −2.5 and AI scores just below 0.2, suggesting that small perturbations in threshold or image quality could shift them across the decision boundary. The predominance of osteopenia among false-positive cases suggests that the AI model may detect a broader spectrum of reduced bone mass rather than exclusively identifying DXA-defined osteoporosis. In a screening context, this tendency may be acceptable because osteopenic individuals can still benefit from further risk assessment and preventive management. However, the model should be regarded as a referral-oriented screening tool that identifies individuals at increased risk of low bone mass, rather than as a diagnostic substitute for DXA-defined osteoporosis. Importantly, subgroup analyses revealed a pronounced diagnostic bias based on baseline BMD category. The model performed strongly in individuals with established DXA-defined osteoporosis, showing high accuracy and sensitivity in this subgroup. However, discrimination within the non-osteoporotic range, particularly among participants with osteopenia, was limited. Because the subgroup analysis was based on the binary definition of DXA-defined osteoporosis, sensitivity and F1 score were not informative for normal and osteopenia groups; in this setting, accuracy provides a more interpretable metric for exploratory comparison across BMD categories. This pattern suggests that the model is better suited to identifying established osteoporosis than to distinguishing subtler gradations of low bone mass on portable CXRs. Such subgroup disparities underscore the importance of dataset balancing, tailored threshold strategies, and potentially dedicated training for early-stage disease detection when deploying AI screening tools in heterogeneous community populations. Sex-stratified analyses provided additional context. Women, who constituted the majority of the cohort, exhibited high sensitivity but lower specificity due to frequent osteopenia-related FPs. Apparent perfect performance in men should be interpreted with caution given the small sample size. Similar sex-related performance trends were reported by Koo et al., indicating that sex-specific evaluation is essential for the appropriate implementation of CXR-based osteoporosis AI (14). From a public health perspective, integrating AI-based osteoporosis screening into existing portable CXR programs could substantially expand access to early detection among frail, homebound, or socioeconomically disadvantaged populations. The KNTA’s mobile service already reaches older adults in community and welfare-center settings where DXA infrastructure is rarely available. In such contexts, a test with high sensitivity and moderate specificity is preferable: FPs primarily incur the cost of confirmatory DXA, whereas false negatives represent truly missed opportunities for fracture prevention. Our findings suggest that at a threshold of 0.2, PROS® CXR: OSTEO offers a reasonable balance for this use case, functioning as a low-cost add-on without additional radiation exposure that is already being performed for another indication. A recent meta-analysis of deep learning models using plain radiographs reported pooled AUC, sensitivity, and specificity of 0.88, 0.81, and 0.87, respectively, supporting the overall feasibility of such opportunistic X-ray-based screening approaches (15). This study has several limitations. First, this was a feasibility-based pilot study with a small sample size (n=52), conducted at a single community screening site in a predominantly female and very elderly population. These factors limit the precision of diagnostic estimates and the generalizability of the findings to other settings, age groups, ethnicities, and sex-balanced populations. Therefore, sex-specific performance should be interpreted descriptively, and the apparently perfect performance in the male subgroup should not be considered evidence of robust model performance in men. Second, only individuals who agreed to undergo DXA at a nearby hospital were included, introducing potential selection bias toward more health-conscious or mobile participants. Third, we evaluated a single commercial AI system without comparison to human readers, other algorithms, or multimodal risk scores. In addition, we did not systematically collect or adjust for clinical risk factors that may influence BMD or radiographic appearance, including prior fracture, glucocorticoid exposure, osteoporosis treatment, calcium or vitamin D supplementation, physical activity, or comorbidities. Future studies should integrate these clinical variables with AI-derived imaging scores to develop a more comprehensive risk-stratification model. Recent multimodal approaches combining X-ray images with clinical data further support the potential value of integrating imaging-derived features with clinical variables for osteoporosis prediction (16). Fourth, although all portable CXRs were acquired using the same device and standardized protocol within a single screening program and were considered diagnostically acceptable for tuberculosis screening, this does not necessarily ensure optimal image quality for AI-based osteoporosis risk assessment. We did not systematically assess image quality, positioning variability, exposure variability, motion, rotation, or radiographic artifacts specifically for AI-based osteoporosis analysis. These factors may have affected AI-derived risk scores and may have contributed to the reduced specificity observed in the portable imaging setting. In addition, the commercial AI system was used as a black-box index test without modification, retraining, or recalibration. Although this reflects real-world implementation, it limits interpretation of the model’s internal decision process, feature attribution, and error patterns. Recent work on CXR-based osteoporosis screening has also emphasized that explainability should be evaluated alongside predictive performance, particularly when AI models are intended for clinical deployment (17). Future studies should include structured image-quality assessment, artifact grading, and portable CXR-specific external validation or recalibration. Finally, because both threshold selection and performance assessment were conducted in the same dataset, the 0.2 operating point should be considered exploratory and requires external validation in a larger, prospectively collected independent portable CXR cohort to reduce the risk of overfitting. Despite these limitations, our study provides initial evidence that a deep learning model developed on conventional hospital-based CXRs can retain clinically meaningful performance when applied to portable CXRs acquired in community screening programs. Future work should include multicenter prospective trials with larger and more diverse populations, explicit analysis of image quality and artifact effects, and exploration of combined strategies that integrate AI-derived risk scores with established clinical fracture risk tools. This direction is consistent with recent reviews suggesting that AI-assisted opportunistic screening using CT or X-ray radiography may help address gaps in osteoporosis detection, while also requiring rigorous validation and cautious clinical implementation (18). Such efforts may help refine portable CXR-based opportunistic screening into a robust, scalable pathway for closing gaps in osteoporosis detection and treatment in underserved communities. Conclusions In this community-based pilot study, a deep learning-based osteoporosis screening model originally developed on fixed digital CXRs maintained meaningful performance when applied to portable CXR acquired during mobile tuberculosis screening. At a threshold of 0.2, the model achieved high sensitivity with moderate specificity, supporting its potential as an opportunistic screening tool for older adults without additional radiation exposure. However, its ability to distinguish early-stage low bone mass in individuals with normal bone density or osteopenia was limited. Given the small and sex-imbalanced cohort, lack of systematic image-quality assessment, and absence of external validation, these findings should be interpreted as preliminary and hypothesis-generating. Larger, more balanced, prospective multicenter studies are needed to validate these findings and determine whether portable CXR-based AI screening can help expand osteoporosis detection in underserved populations. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0668/prf