人工智慧輔助胸部X光骨質疏鬆症篩檢：以台灣實際族群資料探討臨床可行性與診斷公平性

Abstract

骨質疏鬆症為全球重要的公共衛生議題，影響超過 2 億人口，並導致每年約 890 萬件脆弱性骨折，其後續併發症及死亡風險顯著增加。雖然世界各國皆建議以雙能量X光吸收儀(Dual-energy X-ray Absorptiometry, DXA)進行診斷，但現行指引主要針對停經後女性及高齡族群，導致年輕成人、男性及體重正常或偏高者常被排除於常規篩檢之外，使得潛在高風險者無法及早發現與介入。 本研究旨在評估人工智慧(Artificial Intelligence, AI)輔助胸部X光(Chest Radiograph, CXR)於機會性骨質疏鬆篩檢之可行性與臨床效益，並檢驗模型於不同性別、年齡與體重指數(Body Mass Index, BMI)亞群中的表現公平性。研究採回溯性橫斷設計，資料來源為 2012 年至 2023 年間於桃園某區域教學醫院健檢中心接受自費健康檢查之成人共 2,384 名(女性 57.7%、男性 42.3%、平均年齡 43.6 歲)。研究納入同時具備可判讀胸部X光影像及腰椎 DXA 檢測之受試者。AI 模型(VeriOsteo™ OP)為一深度學習系統，利用胸部X光影像自動預測腰椎骨密度(Bone Mineral Density, BMD)及骨質狀態，並與 DXA 結果對照。骨鬆定義依世界衛生組織標準(T-score ≤ –2.5)，50 歲以下者則依國際臨床骨密度學會(ISCD)標準以 Z-score ≤ –2.0 判定為「低於同齡平均值」。 結果顯示，AI 模型標示異常骨密度之比例為 10.7%，而 DXA 確診骨質異常者為 3.9%。在 12 個性別、年齡與 BMI 組合之交叉分層分析中，AI 模型維持穩定高效能，其受試者操作特徵曲線下面積(Area Under the Curve, AUC)介於 0.93–1.00；靈敏度介於 83–100%，陰性預測值(Negative Predictive Value, NPV)為 97–100%，顯示模型具良好偵測與排除能力。模型校準分析顯示預測機率與實際觀察值高度一致(Brier score 0.051；截距 0.068；斜率 1.089)，代表模型預測風險具臨床可解釋性。決策曲線分析(Decision Curve Analysis, DCA)顯示，在臨床具意義之門檻區間(5–30%)內，AI 模型的淨效益(Net Benefit)高於「全部檢測」與「皆不檢測」策略，特別是女性及低 BMI 族群(BMI < 23 kg/m²)，顯示其可作為選擇性轉介 DXA 檢查之輔助工具。相對而言，男性特別是 BMI ≥23 者之淨效益有限，提示其應慎重評估臨床應用價值。 進一步標準化預測值顯示，假設盛行率為女性 10% 與 25%、男性 7% 的情境下，模型於各亞群皆維持高陰性預測值(NPV 96–100%)與低負似然比(LR⁻ 0.06–0.13)，具有高排除骨鬆的可靠性。以女性 ≥50 歲族群為例，若假設盛行率為 25%，每 100 名 AI 預測陽性者約可發現 67–77 例 DXA 確診個案，相當於每 1.3–1.5 次 DXA 即可偵測一名骨鬆患者，顯示其可有效提升檢測效率並降低不必要的 DXA 使用率。 綜合而言，本研究證實 AI 輔助胸部X光分析可作為一項具可行性、可擴展性且具公平性的骨質疏鬆機會性篩檢工具，特別適用於現行指引未涵蓋的族群。此方法可降低設備與人力限制、提升早期偵測率，並補足傳統篩檢的「診斷落差(care gap)」。未來建議進行前瞻性及成本效益研究，評估其於社區及多族群中的推廣潛力，作為數位健康促進與骨鬆防治策略的重要輔助。

Osteoporosis is a major global public health issue, affecting over 200 million people and leading to approximately 8.9 million fragility fractures each year, with substantially increased risks of morbidity and mortality. Although dual-energy X-ray absorptiometry (DXA) remains the gold standard for diagnosis, current screening guidelines mainly target postmenopausal women and older adults. Consequently, younger individuals, men, and those with normal or high body mass index (BMI) are often excluded from routine screening, leaving many high-risk individuals undetected. This study aimed to evaluate the feasibility and clinical utility of an artificial intelligence (AI)-assisted chest radiograph (CXR) model for opportunistic osteoporosis screening, and to examine its diagnostic equity across sex, age, and BMI subgroups. We conducted a retrospective cross-sectional study involving 2,384 adults (57.7% women; mean age, 43.6 years) who underwent self-paid health check-ups at a regional teaching hospital in northern Taiwan between 2012 and 2023. Eligible participants had both evaluable chest X-rays and lumbar spine DXA results. The AI model (VeriOsteo™ OP) is a deep learning system that automatically estimates lumbar bone mineral density (BMD) and bone status from chest radiographs, with DXA measurements serving as the reference standard. Osteoporosis was defined as T-score ≤ –2.5 for participants aged ≥50 years (WHO criteria) and Z-score ≤ –2.0 for those aged <50 years (ISCD definition of “low BMD for age”). The AI model identified 10.7% of participants as having suspected abnormal BMD, while 3.9% were DXA-confirmed abnormal BMD cases. Across 12 sex–age–BMI subgroups, model performance remained robust, with area under the receiver operating characteristic curve (AUC) values ranging from 0.93 to 1.00, sensitivity from 83% to 100%, and negative predictive value (NPV) from 97% to 100%. Calibration analysis showed excellent agreement between predicted and observed probabilities (Brier score = 0.051; intercept = 0.068; slope = 1.089), indicating good reliability of risk estimation. Decision curve analysis demonstrated that within the clinically relevant threshold range (5–30%), the AI model yielded higher net benefit than both “screen-all” and “screen-none” strategies, particularly among women and individuals with BMI <23 kg/m², supporting its potential as a selective triage tool for DXA referral. In contrast, net benefit was limited in men with BMI ≥23 kg/m², suggesting more cautious consideration for clinical implementation in this group. Standardized predictive value analysis further showed that under assumed osteoporosis prevalences of 10% and 25% for women and 7% for men, the model maintained high NPV (96–100%) and low negative likelihood ratios (LR⁻ 0.06–0.13), ensuring reliable rule-out capability. Among women aged ≥50 years with a 25% assumed prevalence, the model detected approximately 67–77 true osteoporosis cases per 100 AI-positive predictions, corresponding to 1.3–1.5 DXA scans per detected case, demonstrating efficiency in reducing unnecessary DXA utilization. In conclusion, this study demonstrates that AI-assisted chest radiograph analysis offers a feasible, scalable, and equitable approach for opportunistic osteoporosis screening, especially for populations not covered by current guideline-based recommendations. By leveraging existing imaging resources, this method can enhance early detection, optimize resource allocation, and help bridge the “osteoporosis care gap.” Future research should focus on prospective validation and cost-effectiveness analyses to evaluate its implementation potential in community and multi-ethnic settings as part of digital health–enabled osteoporosis prevention strategies.