第一部分

原发性甲状旁腺功能亢进症不典型腺瘤与其他病理类型的临床及遗传学特征比较

背景和目的

原发性甲状旁腺功能亢进症（primary hyperparathyroidism，PHPT）是由甲状旁腺原发病变引起甲状旁腺素（parathyroid hormone，PTH）自主分泌过多所致的一种代谢性骨病，其病理类型以良性的甲状旁腺腺瘤（parathyroid adenoma，PA）和甲状旁腺增生（parathyroid hyperplasia，PH）为主，而甲状旁腺不典型腺瘤（atypical parathyroid adenoma，APA）和甲状旁腺癌（parathyroid carcinoma，PC）罕见。APA是较晚被提出的一种恶性潜能未定的病理类型，其临床特征和遗传学特点尚不明确，治疗和管理仍存在争议。本研究拟在单中心大样本PHPT患者队列中，着重探究APA患者在临床特点及预后方面与其他病理类型患者的异同，并分析PHPT相关致病基因胚系变异与临床表型之间的关系，为APA患者的治疗和管理指南共识提供更充足的证据。

对象和方法

纳入2004年1月至2024年7月于本中心内分泌科就诊的PHPT诊断明确且有明确术后病理诊断的患者1012例。回顾性收集患者临床资料，比较良性病变（包括PA和PH）、APA和PC三组患者的临床特点，采用生存分析评价三组的预后并评估APA和PC组术后复发的影响因素。对同意进行基因检测的369例不同时期患者应用不同测序方法，进行PHPT相关致病基因变异筛查，对罕见变异进行ACMG致病性分类。分别将APA和PC患者分为携带和未携带致病性/可能致病性（P/LP）变异的亚组，对2个亚组的临床特征和预后进行比较。

结果

本研究纳入良性病变患者808例（687例PA和121例PH），APA患者104例，PC患者100例：

1. 一般情况及临床表现： APA组诊断年龄48.1±17.1岁，介于良性病变和PC组之间；男性占28.8%，与良性病变组相近并显著低于PC组（56.0%）。APA组临床表现整体介于良恶性两组之间：APA组的骨骼受累（63.5%）、肾结石/肾钙化（52.9%）比例与PC组相近、高于良性病变组；消化道症状（35.6%）和高钙危象（26.0%）比例更接近良性病变组、低于PC组。

2. 实验室指标：APA组的血校正钙、游离钙和PTH水平分别为3.14±0.54 mmol/L、1.50±0.19 mmol/L和578.3 (231.5, 1301.1) pg/mL，均显著高于良性病变组而显著低于PC组（P＜0.01）。APA组低磷血症和高钙尿症比例分别为66.0%和63.6%，介于良恶性两组之间。APA组的血清β-CTX为1.25 (0.81, 2.25)ng/mL、血ALP水平为正常上限的1.48 (0.98, 3.11)倍，与PC组相近，均显著高于良性病变组（P＜0.01）。APA组的肾功能不全（eGFR＜60 mL/min/1.73m²）比例与PC组相近，均显著高于良性病变组（APA: 16.0% (16/100) 和 PC: 17.9% (17/95) vs. 良性病变: 7.9% (61/768)，P值分别为0.017和0.012）。血细胞计数衍生的炎症指标方面，PC组的中性粒细胞-淋巴细胞比值（NLR）、衍生NLR（dNLR）、血小板-淋巴细胞比值（PLR）和系统性免疫炎症指数（SII）均显著高于良性病变组（P＜0.01）；而APA组上述炎症指标与良性病变组相近，其中NLR、dNLR和SII显著低于PC组（P＜0.01）；PC组的淋巴细胞-单核细胞比值（LMR）显著低于良性病变组，APA组与二者均无显著差异。

3. 肿瘤情况及随访：良性病变组多腺体病变比例（9.8%）略高于APA和PC组（5.8%和5.0%）。APA组与PC组的中位肿瘤最大径均为3.0 cm，显著大于良性病变组（1.8cm，P＜0.001）。APA组（11.3%）与良性病变组（16.0%）的复发/未缓解比例无显著差异，均明显低于PC 组（77.8%，P＜0.001）。PC组的中位无复发时间为40.0个月（95%CI: 28.3-51.7），复发风险显著高于APA组（HR=10.24，95%CI: 6.03-17.36）和良性病变组（HR=16.01，95%CI: 8.0-1-32.02）（P＜0.001），而后两组无显著差异。PC、APA、良性病变组的5年无复发生存率依次为33.61%、93.02%、94.21%。多因素Cox回归未发现与APA或PC患者术后复发显著相关的因素。

4. PHPT相关致病基因胚系变异筛查（n=369）：良性病变、APA和PC组的P/LP变异检出率分别为7.7% (22/284)、16.7% (5/30)和27.3% (15/55)。APA和PC患者中最常发生P/LP变异的是CDC73基因，分别占60.0%和80.0%，其次为GCM2基因，而MEN1基因最少见。良性病变组则大多为MEN1基因P/LP变异（86.4%）。APA患者中变异组较非变异组的诊断年龄更早（P=0.014）、血PTH水平更高（P=0.042）、多腺体受累更多见（P=0.023）。APA和PC患者中，变异组的术后复发和未缓解比例均高于非变异组，但未达显著水平。

结论

在本研究队列中：

1. APA 患者的临床和生化特征较甲状旁腺良性病变（PA和PH）患者严重，但较PC患者更轻。

2. APA患者的术后复发和不缓解比例及复发时间与良性病变患者相近，两组的预后均明显优于PC患者。

3. APA患者PHPT相关致病基因的胚系P/LP变异检出率介于良性病变和PC之间，APA和PC患者中CDC73基因P/LP变异最常见，而良性病变患者中MEN1基因P/LP变异最常见。

4. 携带PHPT致病基因胚系P/LP变异的APA患者可能发病更早、更可能出现多腺体受累和术后复发/不缓解，因此需要更积极和更长期的随访。

第二部分

基于机器学习算法构建甲状旁腺癌的术前诊断模型

背景和目的

甲状旁腺癌（parathyroid carcinoma，PC）是原发性甲状旁腺功能亢进症（primary hyperparathyroidism，PHPT）中最少见的病理类型，是PHPT患者手术后不缓解、复发及致死的主要原因之一。良恶性病理类型的PHPT患者的手术方式、预后和随访策略截然不同，但其临床表现、生化指标和影像表现存在重叠，现仍缺乏有效和可靠的术前鉴别诊断方法。通过单一指标粗略地评判会带来较高的误诊和漏诊风险，因此需要一个综合多种术前临床指标的工具来实现PC的术前早期识别。本研究拟基于单中心大样本PHPT患者队列，运用多种机器学习算法构建PC的术前诊断模型，确定最佳模型，进而开发应用工具，辅助临床医师在术前识别PC高风险患者，制定个体化的治疗和随访方案。

对象和方法

纳入1987年12月至2024年7月于本中心内分泌科就诊，PHPT诊断明确并有明确术后病理诊断的患者1089例。回顾性收集患者临床资料并进行数据预处理，采用LASSO回归及逐步Logistic回归筛选关键变量用于构建模型。将总体数据按7: 3随机分为训练集和验证集，在训练集中使用9种机器学习算法构建模型，包括逻辑回归（LR）、k近邻（KNN）、支持向量机（SVM）、随机森林（RF）、梯度提升机（GBM）、轻量级梯度提升机（LightGBM）、自适应增强（AdaBoost）、极端梯度提升（XGBoost）和神经网络（NN）。在验证集中评价模型，区分能力评价采用ROC曲线下面积（AUC）、混淆矩阵，校准能力评价采用校准曲线，临床收益评价采用决策曲线分析（DCA）。对最佳模型进行可视化解释和网页应用部署。

结果

1. 基线资料：纳入122例PC和967例良性病变（包括甲状旁腺腺瘤和增生）患者，随机分为训练集（86例PC和677例良性病变）和验证集（36例PC和290例良性病变）。训练集和验证集的各变量均无显著差异。训练集和验证集中，腺癌组的男性比例、靶器官受累发生率均显著高于良性病变组；腺癌组血校正钙、血甲状旁腺素（PTH）、24小时尿钙（24hUCa）和血碱性磷酸酶（ALP）显著更高，血磷和基于肌酐的估算肾小球滤过率（eGFR）显著更低。

2. 变量筛选：LASSO回归筛选出13个核心变量，进而用向前逐步Logistic回归筛选出8个关键变量：男性、骨骼受累、肾结石/肾钙化、肿瘤最大径、血校正钙、血磷、血PTH和24hUCa，最终纳入模型构建。

3. 模型构建和评价：用上述9种算法构建模型，验证集中DeLong检验提示LightGBM（AUC=0.883）、RF（AUC=0.882）、GBM（AUC=0.878）、LR（AUC=0.876）、SVM（AUC=0.875）、XGBoost（AUC=0.874）和NN（AUC=0.874）模型的AUC两两之间并无统计学差异，而这7个模型的AUC均显著高于KNN（AUC=0.795）和AdaBoost（AUC=0.793）模型。SVM、LightGBM和LR模型的准确度、敏感度、特异度、精确度、阴性预测值和F1分数都较高，模型性能较为平衡。LR和NN的校准曲线最接近参考线。验证集中SVM和LR模型在大多数阈值范围内（0.1-1.0）保持较高净收益并且优于其他模型。综合考虑LR为最佳模型。

4. 模型解释和应用：采用列线图和Sharpley加性解释可视化解释LR模型，得出模型中的变量对诊断的贡献由高到低依次为血钙、性别、24hUCa、血磷、血PTH、骨骼受累、肾结石/肾钙化和肿瘤最大径。该模型已部署为网页在线风险计算器（https://pc-risk-calculator.shinyapps.io/Parathyroid_cancer_risk_calculator/）。

结论

在本研究队列中：

1. 诊断PC的独立危险因素有以下8个，按对诊断模型预测结局的贡献由大到小排序：血校正钙水平高、男性、24hUCa水平高、血磷水平低、血PTH水平高、有骨骼受累、有肾结石/肾钙化及肿瘤直径较大。

2. 基于本中心大样本PHPT队列，LR模型为9种模型中的最优模型，能较有效且可靠地在PHPT患者中术前早期识别PC高风险患者，实现个体化医疗决策，但还需进行外部验证以确保模型在不同人群和临床环境中的稳定性和可靠性。

第三部分

甲状旁腺癌的MGMT基因启动子甲基化和蛋白表达水平研究

背景和目的

甲状旁腺癌（parathyroid carcinoma，PC）的发病率近年来呈上升趋势，相关分子机制研究持续深入，但临床治疗手段仍相对有限，由于PC罕见且分子发病机制复杂，现有研究仅在少数病例中探索过某些可能有效的抗肿瘤药物。近年有小样本发现表观遗传学改变可能是PC的分子病因之一。O⁶-甲基鸟嘌呤-DNA甲基转移酶（O⁶-methylguanine-DNA methyltransferase，MGMT）基因启动子甲基化及其蛋白表达下降与多种恶性肿瘤的发生发展及预后密切相关，但其在PC中的作用尚缺乏研究。本研究拟检测PC患者癌组织中MGMT基因启动子甲基化及其蛋白表达水平，分析其与PC临床特征和预后的相关性，探索MGMT作为新的PC分子病理标志物的潜在可能，也为相关抗肿瘤药物用于PC的治疗提供理论基础。

对象和方法

纳入2008年至2023年于本中心内分泌科诊治的36例PC患者，并按就诊时期、年龄和性别以2: 1匹配甲状旁腺腺瘤（parathyroid adenoma，PA）患者18例，回顾性收集患者的临床资料和甲状旁腺肿瘤组织石蜡切片。采用焦磷酸测序定量检测组织MGMT基因启动子区12个CpG位点（72-83位点）的甲基化水平，采用免疫组化染色半定量检测MGMT蛋白表达水平，用软件分析染色阳性细胞占比和染色强度（OD值）。对比PC和PA组的MGMT启动子甲基化及其蛋白表达水平，比较PC患者不同甲基化水平和不同蛋白表达水平亚组的临床特点和预后。分析MGMT启动子甲基化及其蛋白表达水平的相关性，比较单一和联合指标诊断PC的区分能力。

结果

共纳入36例PC患者的37例癌组织标本（其中1例患者的颈部病灶和肺转移病灶被同时纳入），共28例颈部原发或复发病灶、8例肺转移、1例肝转移标本；匹配18例PA患者的肿瘤组织标本。

1. 基线特征：与PA组相比，PC组靶器官受累比例显著更高、生化改变更严重、肿瘤更大。PC组患者手术后复发/不缓解比例显著高于PA组（91.4% vs. 6.2%，P＜0.001）。

2. MGMT基因启动子甲基化：PC组MGMT基因启动子CpG 72-83各位点甲基化程度存在异质性，其中6个位点（CpG 73-75、77、81-82）的甲基化水平以及CpG 72-83平均甲基化水平（20.3±9.6% vs. 15.2±3.2%，P=0.034）显著高于PA组。CpG 72-83平均甲基化水平区分PC和PA的截断值为18.1%（AUC=0.691，P=0.023），前述6个CpG位点的甲基化水平的AUC值高于CpG 72-83平均甲基化水平。将PC患者分为：低甲基化组（CpG 72-83各位点的甲基化水平均低于相应的最佳截断值，n = 6，16.7 %）、高甲基化组（CpG 72-83各位点的甲基化水平均高于相应的最佳截断值，n = 6，16.7%）和临界组（其余患者，n = 24，66.7%），三组的临床表现、骨密度、生化指标和肿瘤大小均无显著差异。低甲基化组、临界组和高甲基化组的术后复发/未缓解比例依次升高（83.3% vs. 91.3% vs. 100.0%），复发时间呈缩短趋势（89.5 (18.0, 155.8) vs. 31.0 (16.0, 60.0) vs. 24.0 (16.0, 35.0) 个月），但差异均未达显著性水平。

3. MGMT蛋白表达：PC组的MGMT染色阳性细胞占比和OD值均显著低于PA组，分别为23.0% (1.9%, 65.0%) vs. 77.6% (57.8%, 90.9%)（P＜0.001）和0.066 (0.044, 0.092) vs. 0.079 (0.069, 0.090) （P=0.002）。将PC患者按MGMT阳性细胞占比（以50%为界）分为低表达组（n=25，69.4%）和高表达组（n=10，27.8%）。两组患者的临床表现、骨密度、生化指标和肿瘤大小均无显著差异。低表达组术后复发/未缓解比例（96.0% vs. 80.0%）、转移比例（72.0% vs. 45.5%）高于高表达组，复发时间短于高表达组，分别为29.0 (16.0, 42.3) 个月和42.0 (18.0, 90.8) 个月，但差异均未达显著性水平。

4. MGMT启动子甲基化与蛋白表达的关联：PC组中由低甲基化到高甲基化组， MGMT蛋白染色的中位阳性细胞占比（65.0% vs. 33.0% vs. 19.5%）依次降低，但差异未达显著性水平。Logistic回归显示MGMT蛋白低表达（OR=14.224，95%CI: 2.948-68.643，P＜0.001）和CpG 72-83平均甲基化水平高（OR=1.127，95%CI: 1.007-1.261，P=0.038）均为PC的独立危险因素。CpG 72-83平均甲基化水平为18.1%时诊断PC的AUC=0.691（95%CI: 0.552-0.829，P=0.023），而联合蛋白表达水平后AUC显著升至0.844（95%CI: 0.734-0.955），P＜0.001。

结论

在本组PC及PA病例中：

1. 与PA组织相比，PC组织存在MGMT基因启动子高甲基化及其蛋白表达下降。

2. MGMT启动子高甲基化及其蛋白表达下降的PC患者术后复发/不缓解风险有增高趋势，复发时间可能更短。

3. 判断MGMT启动子甲基化水平的切点为CpG 72-83平均值18.1%，联合甲基化及蛋白表达水平能更有效区分PC和PA。

Part 1

Clinical and Genetic Features of Atypical Adenoma in Primary Hyperparathyroidism Compared with Other Pathological Types

Background and Objective

Primary hyperparathyroidism (PHPT) is a metabolic bone disorder caused by primary lesions of parathyroid glands with excessive autonomous secretion of parathyroid hormone (PTH). The pathological types of PHPT are predominantly benign parathyroid adenoma (PA) and parathyroid hyperplasia (PH), while atypical parathyroid adenoma (APA) and parathyroid carcinoma (PC) are rare. APA is a recently recognized pathological type with uncertain malignant potential, and its clinical and genetic characteristics remain unclear, with controversial treatment and management. This study aims to explore the differences in the clinical features and prognosis of APA patients compared to patients with other pathological types in a large single-center PHPT cohort, and to analyze the relationship between germline variations in PHPT-related genes and clinical phenotypes, providing more evidence for the treatment and management guidelines of APA patients.

Subjects and Methods

We included 1,012 PHPT patients with postoperative pathological diagnosis who visited the endocrinology department of our center between January 2004 and July 2024. Clinical data were retrospectively collected. The clinical features of benign lesions (including PA and PH), APA, and PC groups were compared. Survival analysis was conducted to evaluate the prognosis of the three groups and to assess the influencing factors for postoperative recurrence in the APA and PC groups. For 369 patients who consented to genetic testing at different periods, various sequencing methods were used for screening PHPT-related gene variations. Rare variations were classified for pathogenicity according to ACMG guidelines. APA and PC patients were divided into subgroups with and without pathogenic/likely pathogenic (P/LP) variations, and their clinical features and prognosis were compared.

Results

A total of 808 benign lesion patients (687 PA and 121 PH), 104 APA patients, and 100 PC patients were included.

1. Demographics and Clinical Manifestations: The mean age at diagnosis of APA group was 48.1 ± 17.1 years, which was between the benign lesion and PC groups. The proportion of males of APA group was 28.8%, similar to the benign lesion group and significantly lower than the PC group (56.0%). Clinical manifestations of APA patients were intermediate between benign and malignant groups: the proportions of skeletal involvement (63.5%) and urolithiasis/renal calcification (52.9%) were similar to the PC group and higher than the benign lesion group, while the incidences of gastrointestinal symptoms (35.6%) and hypercalcemic crises (26.0%) of APA patients were closer to the benign lesion group and lower than the PC group.

2. Laboratory Indicators: The serum corrected calcium, serum ionized calcium, and PTH levels in the APA group were 3.14 ± 0.54 mmol/L, 1.50 ± 0.19 mmol/L, and 578.3 (231.5, 1301.1) pg/mL, respectively, significantly higher than the benign lesion group and lower than the PC group (P < 0.01). The proportion of hypophosphatemia and hypercalciuria of APA group were 66.0% and 63.6%, respectively, falling between the benign and malignant groups. The serum β-CTX level of APA group was 1.25 (0.81, 2.25) ng/mL, and serum ALP level was 1.48 (0.98, 3.11) times the normal upper limit, which was close to the PC group and significantly higher than the benign lesion group (P < 0.01). The incidence of renal insufficiency (eGFR < 60 mL/min/1.73m²) in the APA group was resemble to the PC group and significantly higher than the benign lesion group (APA: 16.0% (16/100) and PC: 17.9% (17/95) vs. benign lesion: 7.9% (61/768), P = 0.017 and 0.012, respectively). As for complete blood count-derived inflammatory indicators, the neutrophil-to-lymphocyte ratio (NLR), derived NLR (dNLR), platelet-to-lymphocyte ratio (PLR), and systemic immune-inflammatory index (SII) in the PC group were significantly higher than the benign lesion group (P < 0.01), while the APA group showed similar values to the benign lesion group, with NLR, dNLR, and SII significantly lower than the PC group (P < 0.01). The lymphocyte-to-monocyte ratio (LMR) in the PC group was significantly lower than the benign lesion group, and the LMR of APA group was not different from either of the other groups.

3. Tumor Characteristics and Follow-up: The proportion of multiglandular disease in the benign lesion group (9.8%) was slightly higher than that in the APA and PC groups (5.8% and 5.0%). The median maximum tumor diameters of APA and PC were both 3.0 cm, significantly larger than that of benign lesion (1.8 cm, P < 0.001). The recurrence/persistence rate of PHPT in the APA group (11.3%) was not significantly different from the benign lesion group (16.0%), and both were remarkably lower than the PC group (77.8%, P < 0.001). The median recurrence-free survival (RFS) time of PC patients was 40.0 months (95% CI: 28.3-51.7), with a significantly higher relapse risk compared to the APA group (HR = 10.24, 95% CI: 6.03-17.36) and the benign lesion group (HR = 16.01, 95% CI: 8.01-32.02) (P < 0.001), while the latter two groups had no significant difference. The 5-year RFS rates of PC, APA, and benign lesion groups were 33.61%, 93.02%, and 94.21%, respectively. Multivariate Cox regression analysis identified no significant factors associated with postoperative recurrence in the APA or PC groups.

4. Screening for Germline Variations in PHPT-related Genes (n=369): The rates of P/LP variations in the benign lesion, APA, and PC groups were 7.7% (22/284), 16.7% (5/30), and 27.3% (15/55), respectively. In APA and PC patients, P/LP variants are most frequently detected in the CDC73 gene, accounting for 60.0% and 80.0% of cases respectively, followed by the GCM2 gene, while MEN1 P/LP variants are the least common. While in the benign lesion group, P/LP variants are most frequently detected in the MEN1 gene (86.4%). APA patients carrying P/LP variants had younger age at diagnosis (P = 0.014), higher PTH levels (P = 0.042), and more common multiglandular involvement (P = 0.023), compared to those without P/LP variants. In both the APA and PC groups, the variation group had a higher incidence of postsurgical recurrence/non-remission than the non-variation group, but the differences were not significant.

Conclusion

In this study cohort:

1. The clinical and biochemical characteristics of APA patients are more severe than those of patients with benign parathyroid lesions (PA and PH), but milder than those of PC patients.

2. The incidence of postoperative recurrence/non-remission and the time to relapse in APA patients are similar to those in patients with benign lesions, and both groups have a significantly better prognosis than PC patients.

3. The rate of germline P/LP variants in PHPT-related genes in APA patients falls between that of benign lesions and PC groups. In APA and PC patients, CDC73 is the gene most frequently harboring P/LP variants, whereas in benign lesions group, MEN1 is the most common gene with P/LP variants.

4. APA patients carrying germline P/LP variants in PHPT-related genes may develop the disease earlier and have an increased risk of multiglandular involvement and postoperative recurrence or non-remission, thus requiring more proactive and longer-term follow-up strategies.

Part 2

Machine Learning–Based Preoperative Diagnostic Model for Parathyroid Carcinoma

Background and Objective

Parathyroid carcinoma (PC) is the rarest pathological type of primary hyperparathyroidism (PHPT) and a main cause of persistent disease, recurrence, and mortality following surgery in PHPT patients. Although benign and malignant pathological type of PHPT differ significantly in surgical approach, prognosis, and follow-up strategy, their clinical, biochemical, and imaging features can overlap. A reliable and effective preoperative method for differentiating PC is still lacking. Relying on a single indicator carries a high risk of both misdiagnosis and missed diagnosis. Therefore, a tool integrating multiple preoperative clinical indicators is needed to achieve early diagnosis of PC. This study aims to develop a preoperative diagnostic model for PC using various machine learning algorithms based on a large single-center PHPT cohort, identify the optimal model, and translate it into a web application to help clinicians recognize PC high-risk patients preoperatively and take individualized treatment and follow-up strategies.

Subjects and Methods

A total of 1,089 PHPT patients with definitive postoperative pathology, who visited our center between December 1987 and July 2024, were retrospectively included. After collecting clinical data and performing data preprocessing, LASSO regression and stepwise logistic regression were conducted to identify key variables for model construction. The dataset was randomly split 7: 3 into a training set and a validation set. Nine machine learning algorithms were used to build models in the training set: logistic regression (LR), k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), light gradient boosting machine (LightGBM), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and neural network (NN). In the validation set, each model’s performance was assessed by evaluating discrimination with the area under the ROC curve (AUC) and confusion matrices, evaluating calibration with calibration curves, and evaluating clinical utility with decision curve analysis (DCA). The best-performing model was then visualized and interpreted and deployed as an online application.

Results

1. Baseline Data: A total of 122 PC and 967 benign lesions (adenoma or hyperplasia) patients were included. They were randomly assigned to the training set (86 PC, 677 benign lesions) and the validation set (36 PC, 290 benign lesions). No variable differed significantly between the two sets. In both sets, compared with the benign group, the PC group had a higher proportion of males and higher incidence of target-organ involvement, as well as elevated serum corrected calcium (cCa), parathyroid hormone (PTH), 24-hour urinary calcium (24hUCa), and alkaline phosphatase (ALP), and lower serum phosphate and estimated glomerular filtration rate based on creatinine (eGFR).

2. Variable Selection: LASSO regression identified 13 core variables, and further stepwise logistic regression narrowed these down to 8 key factors: male sex, skeletal involvement, urolithiasis/nephrocalcinosis, maximum tumor diameter, serum cCa, serum phosphate, serum PTH, and 24hUCa. 

3. Model Construction and Evaluation: In the validation set, DeLong’s test showed that the AUCs of LightGBM (0.883), RF (0.882), GBM (0.878), LR (0.876), SVM (0.875), XGBoost (0.874), and NN (0.874) did not differ significantly from one another; however, all these 7 models’ AUCs were higher than KNN (0.795) and AdaBoost (0.793). SVM, LightGBM, and LR had relatively balanced performance regarding accuracy, sensitivity, specificity, precision, negative predictive value, and F1 score. LR and NN showed better calibration, while SVM and LR yielded higher net benefit in DCA within most threshold ranges (0.1–1.0) compared to other models. Overall, LR was considered as the optimal model.

4. Model Interpretation and Application: Nomogram and SHAP (SHapley Additive exPlanations) methods were used to visualize and interpret the LR model. The contribution of each predictor, from greatest to least, was serum cCa, sex, 24hUCa, serum phosphate, serum PTH, skeletal involvement, urolithiasis/nephrocalcinosis, and maximum tumor diameter. The LR model is available as a web-based risk calculator (https://pc-risk-calculator.shinyapps.io/Parathyroid_cancer_risk_calculator/ ).

Conclusion

In this study cohort:

1. Eight independent risk factors for PC, ranked by their contribution to the diagnostic model, are higher serum cCa, male sex, higher 24hUCa, lower serum phosphate, higher serum PTH, skeletal involvement, urolithiasis/nephrocalcinosis, and larger tumor diameter. 

2. Based on a large single-center PHPT cohort, the logistic regression model achieved the best performance among nine machine learning models for early preoperative recognition of patients with high risk for PC, facilitating personalized therapeutic decision-making. Further external validation is required to confirm its stability and reliability in other populations and clinical settings. 

Part 3

Study on MGMT Gene Promoter Methylation and Protein Expression in Parathyroid Carcinoma

Background and Objective

The incidence of parathyroid carcinoma (PC) is rising in recent years, with ongoing research into its molecular mechanisms. However, clinical treatment options remain limited. Due to the rarity of PC and its complex molecular pathogenesis, existing studies have explored potential anti-tumor drugs in only a few cases. Recent studies suggest that epigenetic changes may contribute to the molecular etiology of PC. O⁶-methylguanine-DNA methyltransferase (MGMT) gene promoter methylation and protein expression have been associated with the development and prognosis of various malignancies, but studies on their role in PC remains lacking. This study aims to examine MGMT gene promoter methylation and its protein expression in cancerous tissues from PC patients, assess their correlation with clinical features and prognosis of PC patients, and explore the potential of MGMT as a new molecular pathological marker for PC. The findings may also provide theoretical support for the use of related anti-tumor drugs in treating PC.

Subjects and Methods

A total of 36 PC patients treated at our center between 2008 and 2023 were included. For comparison, 18 parathyroid adenoma (PA) patients were matched to the PC patients based on age, sex, and period of visit in a 1: 2 ratio. Clinical data and paraffin-embedded parathyroid tumor tissue specimens were retrospectively collected. The methylation levels of 12 CpG sites (CpG 72–83) in the MGMT gene promoter were quantitatively measured by pyrosequencing. MGMT protein expression was semi-quantitatively assessed using immunohistochemistry, analyzing the percentage of positive cells and staining intensity (OD values). MGMT promoter methylation and its protein expression levels were compared between the PC and PA groups. The clinical characteristics and prognosis of PC patients with different methylation levels and protein expression subgroups were compared. The association between MGMT promoter methylation and its protein expression was analyzed, and the diagnostic ability of individual and combined markers for distinguishing PC was evaluated.

Results

A total of 37 PC tissue specimens from 36 PC patients were included (one patient provided specimens from both a neck lesion and a lung metastasis), comprising 28 specimens from primary or recurrent neck lesions, 8 from lung metastases, and 1 from a liver metastasis. For comparison, 18 tumor specimens from PA patients were included.

1. Baseline Characteristics: Compared with the PA group, PC patients had a significantly higher rate of target-organ involvement, more severe biochemical disturbances, and larger tumor size. Furthermore, the incidence of postoperative recurrence/non-remission in the PC group was markedly higher than that in the PA group (91.4% vs. 6.2%, P < 0.001).

2. MGMT Gene Promoter Methylation: The methylation levels of CpG 72-83 in the MGMT gene promoter showed heterogeneity in the PC group. Six CpG sites (CpG 73-75, 77, 81-82) and the average methylation level of CpG 72-83 were significantly higher in the PC group than in the PA group (20.3±9.6% vs. 15.2±3.2%, P = 0.034). The threshold for distinguishing PC and PA based on the average methylation level of CpG 72-83 was 18.1% (AUC = 0.691, P = 0.023). The AUC values of the methylation levels of the six CpG sites were higher than that of the average methylation level of CpG 72-83. PC patients were stratified into three subgroups: low-methylation group (methylation levels at each CpG site 72–83 below the respective optimal cut-off values, n = 6, 16.7%), high-methylation group (methylation levels at each CpG site 72–83 above the respective optimal cut-off values, n = 6, 16.7%), and a borderline group (comprising the remaining patients, n = 24, 66.7%). There were no significant differences in clinical features, bone mineral density, biochemical indicators, or tumor size among the three groups. However, the postoperative recurrence/non-remission rates increased from the low-methylation group to the high-methylation group (83.3% vs. 91.3% vs. 100.0%), and the time to relapse tended to decrease (89.5 (18.0, 155.8) vs. 31.0 (16.0, 60.0) vs. 24.0 (16.0, 35.0) months), though the differences were not statistically significant. 

3. MGMT Protein Expression: The percentage of MGMT-positive cells and the OD values were significantly lower in the PC group than in the PA group (23.0% (1.9%, 65.0%) vs. 77.6% (57.8%, 90.9%) (P < 0.001) and 0.066 (0.044, 0.092) vs. 0.079 (0.069, 0.090) (P = 0.002)), respectively. Based on the percentage of MGMT-positive cells (cut-off value = 50%), PC patients were divided into low-expression (n = 25, 69.4%) and high-expression (n = 10, 27.8%) groups. No significant differences in clinical features, bone mineral density, biochemical markers, or tumor size were observed between the two groups. Compared to the high-expression group, the low-expression group had higher postoperative recurrence/non-remission rates (96.0% vs. 80.0%), higher metastasis rates (72.0% vs. 45.5%), and shorter time to relapse (29.0 (16.0, 42.3) vs. 42.0 (18.0, 90.8) months), but the differences were not statistically significant. 

4. Correlation Between MGMT Promoter Methylation and Protein Expression: In the PC group, the median percentage of MGMT-positive cells decreased from the low-methylation to the high-methylation group (65.0% vs. 33.0% vs. 19.5%), but the differences were not statistically significant. Logistic regression showed that low MGMT protein expression (OR = 14.224, 95% CI: 2.948–68.643, P < 0.001) and the average methylation level of CpG 72-83 (OR = 1.127, 95% CI: 1.007–1.261, P = 0.038) were independent risk factors for PC. The AUC for distinguishing PC using the CpG 72-83 average methylation level at 18.1% was 0.691 (95% CI: 0.552–0.829, P = 0.023). After combining protein expression, the AUC increased significantly to 0.844 (95% CI: 0.734–0.955, P < 0.001).

Conclusion

In this cohort of PC and PA cases:

1. Compared to PA tissues, PC tissues exhibit higher MGMT gene promoter methylation and lower MGMT protein expression. 

2. PC patients with high MGMT promoter methylation and low MGMT protein expression probably have increased risk of postoperative recurrence/non-remission and may have shorter time to relapse.

3. A cut-off value for determining MGMT promoter methylation is established as CpG 72-83 average methylation level of 18.1%, and combined this with MGMT protein expression level can more effectively distinguish between PC and PA.