Obsessive-compulsive disorder (OCD) is characterized by time-consuming obsessions and compulsions that cause distress and impairment¹. It can affect people of all ages and has a lifetime prevalence of 1-2%^{2, 3}. The severity of OCD is assessed with the Yale-Brown Obsessive Compulsive Scale (Y-BOCS)^{4, 5}. Despite extensive use of this scale for several decades, there is still uncertainty about what constitutes subclinical, mild, moderate and severe OCD.

To our knowledge, only two previous studies have attempted to calculate Y-BOCS severity benchmarks^{6, 7}, yielding inconsistent results. Both studies were underpowered, as they included a small number of individuals in the lower and higher severity ends of the distribution, and only recruited participants from a single country or single age group.

To provide definitive severity benchmarks for OCD that can be used across the lifespan and different cultures, large multinational samples are required. Empirically supported severity benchmarks would facilitate clinical decision making, trial design, and communication between professionals, the patient community and policy makers.

The OCD Severity Benchmark Consortium collected Y-BOCS data from 5,140 individuals with a lifetime diagnosis of OCD from Sweden, Brazil, South Africa, US and India (47/53% male/female, 21/79% children/adults, age range: 5-82 years). Data were collected as part of various research projects; each of the individual studies was approved by the local ethical review board, and all participants provided written informed consent (or assent if under the age of 18) for participation.

Data from four countries were used for model development (Sweden, N=1,697; Brazil, N=936; South Africa, N=552; US, N=599; total N=3,784). Data from India (N=1,356) were used for external model validation. Experienced clinicians administered the child or adult versions of the Y-BOCS, and the Clinical Global Impression-Severity (CGI-S) scale, which constituted the benchmark measure in this study. The CGI-S is a single-item measure (score range: 1-7) of global disorder severity (in this case, OCD) that synthesizes all available information about the patient, including but not limited to current symptoms, impairment and general function⁸.

An ordinal logistic regression model was trained in 80% of the data from the four countries used for model development (training dataset, N=3,027) and accuracy of the best severity benchmarks was separately evaluated in the remaining 20% of these data (holdout dataset, N=757) and in the external dataset from India. To compensate for the unevenly distributed severity classes during model development, oversampling was performed by drawing 2,500 samples, with replacement, from each severity class.

A large proportion of all participants in the training and holdout datasets were classified as having moderately severe OCD (CGI-S score of 4 or 5; N=2,577, 68.1%). The next most common severity class was mild OCD (CGI-S score of 3; N=580, 15.3%), followed by severe OCD (CGI-S score of 6 or 7; N=408, 10.8%), and subclinical OCD (CGI-S score of 1 or 2; N=219, 5.8%). In the external Indian dataset, moderately severe OCD was most common (N=502, 37.0%), followed by severe OCD (N=352, 26.0%), mild OCD (N=341, 25.1%), and subclinical OCD (N=161, 11.9%).

Spearman’s rho indicated that severity class and Y-BOCS severity correlated moderately to strongly (r=.61, p<0.00001). An ordinal regression model with severity class as the dependent variable and Y-BOCS score as the independent variable was statistically significant (p<0.00001), and the Nagelkerke’s pseudo R² estimate of the model indicated that variation in Y-BOCS severity accounted for 47.9% of the variation in the CGI-S severity classification.

Using the training dataset, the ordinal regression model indicated that subclinical OCD corresponded to scores of 0-13 points on the Y-BOCS, mild OCD to 14-21 points, moderate OCD to 22-29 points, and severe OCD to 30-40 points. These benchmarks classified individuals in the holdout and external datasets with modest accuracy (holdout: 57%, external: 55%). When we allowed the severity levels to overlap three points, accuracy increased to 79% in both datasets. This indicates that roughly half of misclassifications appeared around the breakpoints, which is expected since OCD severity is a dimensional construct⁹.

A Y-BOCS score of 14 points separated clinical from subclinical individuals with excellent sensitivity (holdout: 94%, external: 91%) and adequate specificity (62% and 78%, respectively). The positive predictive value (PPV), or proportion of participants classified as having clinical OCD who truly had clinical OCD, was excellent in both the holdout (98%) and the external (99%) datasets. The negative predictive value (NPV), or proportion of participants classified as having subclinical OCD that truly had subclinical OCD, was lower (40% and 28%, respectively).

Interestingly, 14 is two points lower than the 16 points that are typically used as inclusion criteria for entry in most clinical trials of OCD. To the best of our knowledge, the 16-point cut-off used in clinical trials is arbitrary and could be revised in light of the current findings.

A Y-BOCS score of 30 points separated severe from non-severe OCD with adequate sensitivity (holdout: 70%, external: 82%), good specificity (89% and 84%), a low PPV (43% and 49%), and a high NPV (96% and 96%). Thus, a score of 30 may work best to screen out individuals with severe OCD rather than identifying a pure group above a certain severity level. Therefore, decisions to ration access to certain intensive specialist treatments to individuals with Y-BOCS scores above 30 should be questioned.

Largely consistent classification performance (total accuracy, sensitivity, specificity, PPV and NPV) of the general benchmarks was found across countries, genders and age groups, and overall benchmarks were similar in accuracy to subgroup-derived benchmarks (i.e., benchmarks that were based on only ­subgroups of the training dataset). This indicates that the provided benchmarks are largely invariant across national settings and individuals, and can therefore be used globally and across the lifespan.

In summary, we provide the field with empirically derived Y-BOCS severity benchmarks across the lifespan which will be useful in research and clinical settings (subclinical OCD: 0-13 points; mild OCD: 14-21 points; moderate OCD: 22-29 points; severe OCD: 30-40 points).

However, due to the modest accuracy of the classifications, we caution against the exclusive use of these benchmarks to guide important clinical decisions regarding individual patients, such as offering access to specialist treatment. Other relevant variables should be used, together with Y-BOCS scores, to guide clinical decision making and resource allocation, such as duration of the disorder, time without adequate treatment, psychiatric and somatic comorbidities, family accommodation, socioeconomic circumstances, and personal treatment history.

强迫症 (OCD) 的特点是耗时的强迫观念和强迫行为，会导致痛苦和损害¹。它可以影响所有年龄段的人，终生患病率为 1-2% ^{2, 3}。强迫症的严重程度使用耶鲁-布朗强迫症量表 (Y-BOCS) ^{4, 5}进行评估。尽管这个量表被广泛使用了几十年，但对于亚临床、轻度、中度和重度强迫症的构成仍然存在不确定性。

据我们所知，只有两项先前的研究试图计算 Y-BOCS 严重性基准^6、7，得出的结果不一致。这两项研究的功效都不足，因为它们在分布的较低和较高严重性端都包括少数个体，并且只招募了来自单个国家或单个年龄组的参与者。

为了提供可在整个生命周期和不同文化中使用的 OCD 的明确严重程度基准，需要大型跨国样本。经验支持的严重性基准将促进临床决策、试验设计以及专业人员、患者社区和政策制定者之间的沟通。

OCD 严重性基准联盟从瑞典、巴西、南非、美国和印度（47/53% 男性/女性，21/79% 儿童/成人，年龄范围： 5-82 岁）。作为各种研究项目的一部分收集数据；每项单独的研究都得到了当地伦理审查委员会的批准，所有参与者都提供了参与的书面知情同意书（或同意，如果未满 18 岁）。

来自四个国家的数据用于模型开发（瑞典，N=1,697；巴西，N=936；南非，N=552；美国，N=599；总 N=3,784）。来自印度的数据 (N=1,356) 用于外部模型验证。经验丰富的临床医生管理儿童或成人版本的 Y-BOCS，以及构成本研究基准测量的临床总体印象严重程度 (CGI-S) 量表。CGI-S 是对整体疾病严重程度（在本例中为强迫症）的单项测量（评分范围：1-7），它综合了有关患者的所有可用信息，包括但不限于当前症状、损伤和一般功能⁸ .

在用于模型开发的四个国家的 80% 的数据（训练数据集，N=3,027）中训练了序数逻辑回归模型，并在其余 20% 的数据（保留数据集）中分别评估了最佳严重性基准的准确性, N=757) 和来自印度的外部数据集中。为了弥补模型开发过程中严重性等级分布不均的问题，通过从每个严重性等级中抽取 2,500 个样本进行过采样，并进行替换。

训练和坚持数据集中的所有参与者中有很大一部分被归类为中度重度强迫症（CGI-S 评分为 4 或 5；N=2,577, 68.1%）。下一个最常见的严重程度等级是轻度强迫症（CGI-S 评分为 3；N=580，15.3%），其次是重度强迫症（CGI-S 评分为 6 或 7；N=408，10.8%）和亚​​临床强迫症（CGI-S 评分为 1 或 2；N=219, 5.8%）。在外部印度数据集中，中重度强迫症最常见（N=502, 37.0%），其次是重度强迫症（N=352, 26.0%）、轻度强迫症（N=341, 25.1%）和亚临床强迫症（N =161, 11.9%)。

Spearman 的 rho 表明严重程度等级和 Y-BOCS 严重程度中度至强相关（r=.61，p<0.00001）。以严重程度等级为因变量、Y-BOCS 评分为自变量的有序回归模型具有统计学意义（p<0.00001），模型的 Nagelkerke 伪 R ²估计表明 Y-BOCS 严重程度的变化占 47.9 CGI-S 严重性分类中变化的百分比。

使用训练数据集，序数回归模型表明，亚临床强迫症对应于 Y-BOCS 得分 0-13 分，轻度强迫症对应 14-21 分，中度强迫症对应 22-29 分，重度强迫症对应 30-40 分点。这些基准以适度的准确度对坚持和外部数据集中的个人进行分类（坚持：57%，外部：55%）。当我们允许严重性级别重叠三个点时，两个数据集中的准确度都提高到了 79%。这表明大约一半的错误分类出现在断点周围，这是预期的，因为 OCD 严重程度是一个维度结构⁹。

14 分的 Y-BOCS 评分将临床个体与亚临床个体区分开来，具有出色的敏感性（坚持：94%，外部：91%）和足够的特异性（分别为 62% 和 78%）。阳性预测值 (PPV) 或被归类为患有临床强迫症但真正患有临床强迫症的参与者的比例在保留 (98%) 和外部 (99%) 数据集中都非常出色。阴性预测值 (NPV) 或被归类为亚临床强迫症的参与者真正患有亚临床强迫症的比例较低（分别为 40% 和 28%）。

有趣的是，14 分比大多数强迫症临床试验中通常用作纳入标准的 16 分低 2 分。据我们所知，临床试验中使用的 16 点临界值是任意的，可以根据目前的发现进行修改。

30 分的 Y-BOCS 评分将严重与非严重强迫症区分开来，具有足够的敏感性（保留：70%，外部：82%），良好的特异性（89% 和 84%），低 PPV（43% 和 49%） ，以及高 NPV（96% 和 96%）。因此，30 分可能最适合筛选出患有严重强迫症的个体，而不是识别出某个严重程度以上的纯群体。因此，应该质疑对 Y-BOCS 分数高于 30 的个人分配某些强化专科治疗的决定。

在不同国家、性别和年龄组中发现一般基准的分类性能（总准确度、敏感性、特异性、PPV 和 NPV）大体一致，总体基准在准确度上与亚组衍生基准（即基于仅训练数据集的子组）。这表明所提供的基准在不同国家环境和个人之间基本上是不变的，因此可以在全球和整个生命周期中使用。

总之，我们为该领域提供了在整个生命周期中根据经验得出的 Y-BOCS 严重程度基准，这将在研究和临床环境中有用（亚临床强迫症：0-13 分；轻度强迫症：14-21 分；中度强迫症：22-29分；重度强迫症：30-40 分）。

然而，由于分类的准确性不高，我们告诫不要只使用这些基准来指导有关个体患者的重要临床决策，例如提供专科治疗的机会。其他相关变量应与 Y-BOCS 评分一起用于指导临床决策和资源分配，例如疾病的持续时间、未得到充分治疗的时间、精神和躯体合并症、家庭住宿、社会经济环境和个人治疗史.