人工智能器械的性能评价难在哪

结论：难点不在算法，在用什么数据来评价

基于机器学习的器械，性能评价的困难与传统器械不同。传统器械的性能可以用物理量直接测量——测十次结果应当一致，与谁来测、在哪测关系不大。

机器学习模型不同： 它的性能取决于输入数据的分布。在一批数据上表现很好的模型，换一批数据可能明显下降。

所以评价的核心问题是：用什么数据来评价，这批数据能否代表实际使用中会遇到的情况。 数据选得不对，测出来的性能指标再好看也说明不了实际表现。

几个特有的难点

数据依赖性。 性能与测试数据的分布强相关。

泛化能力难以确认。 在训练数据的分布之外表现如何，难以完全预知。

可解释性有限。 模型给出某个结论的原因往往难以说清。

性能可能随时间变化。 实际使用中数据分布发生变化（人群变化、设备更新、操作习惯变化）时，性能可能下降。

失效模式不直观。 传统器械的失效有明确的物理原因，模型的错误输出往往找不到直接原因。

边界情况难以穷举。 罕见但重要的情况可能在数据中样本很少。

这几点决定了评价不能只看一个准确率数字，需要更完整的证据组织。

数据集的要求

评价用的数据集是核心，要求包括：

「独立性」是最基本也最容易出问题的要求。 测试数据如果与训练数据有重叠（哪怕是同一个个体的不同样本），测出的性能会明显偏高，不代表真实能力。

「代表性」要具体说明。 数据来自几个来源、覆盖什么人群分布、什么设备型号、什么采集条件。只说「数据集包含多少样本」而不说构成，代表性就无从判断。

参考标准的确定

评价需要「正确答案」作为对照，这个参考标准怎么来，直接影响结论：

参考标准的来源。 专家判读、其他检查方法、随访结果、病理结果。

参考标准本身的可靠性。 专家判读也会有分歧和误差。

多人判读的处理。 判读不一致时如何确定参考标准。

判读者的资质与盲法。 判读时是否知道模型的结论。

参考标准的局限。 参考标准本身可能不完美，这会影响性能指标的解读。

「判读者是否盲法」是评价质量的关键。 如果判读者看到了模型的输出再做判断，参考标准就被污染了。

指标的选择

单一指标往往不足以描述性能：

准确率。 直观但在类别不平衡时具有误导性。

灵敏度与特异度。 分别反映对阳性和阴性的识别能力。

阳性预测值与阴性预测值。 与实际患病率相关，更贴近临床解读。

受试者工作特征相关指标。 反映不同阈值下的综合表现。

置信区间。 反映估计的不确定性。

分层分析。 在不同人群、不同设备、不同条件下的表现。

「分层分析」是评价中很有价值的部分。 整体性能好，但在某个亚组（比如某个年龄段、某种设备采集的数据）上明显差，这个信息对安全很重要，而总体指标掩盖了它。

泛化能力的考察

外部验证。 用与开发数据完全不同来源的数据验证。

多中心数据。 覆盖不同机构、不同设备。

时间上的外部性。 用不同时间段的数据验证。

边界情况的表现。 罕见情况、图像质量差、非典型表现。

对抗性输入。 异常输入时的行为，是否给出明显错误的高置信度结论。

「外部验证」比内部测试更能说明泛化能力，因为内部测试的数据即使独立，也来自同样的采集环境和人群。

使用中的监测

性能可能随时间变化，所以上市后的监测是必要的：

输入数据分布的监测。 实际使用中的数据是否偏离了开发时的分布。

性能的持续评估。 定期用新数据评估。

用户反馈的收集。 错误案例的收集与分析。

更新的管理。 模型更新时的验证和申报。

「输入数据分布的偏移」是这类产品特有的风险。 设备更新、操作规程变化、人群构成变化，都可能使实际数据偏离开发时的分布，而模型性能随之下降，这个过程可能是渐进且不易察觉的。

资料的组织

提交的性能评价资料建议包含：

数据集的详细描述（来源、构成、分布）；独立性的保证措施；参考标准的建立方法；评价方案与预设的指标和判据；总体性能结果与置信区间；分层分析结果；边界情况的分析；以及上市后的监测计划。

「预设的指标和判据」要在评价前确定并记录。 做完评价再选择表现好的指标来报告，是评价方法上的问题。

与临床评价的关系

性能评价与临床评价是两个层次：

性能评价。 模型在测试数据上的技术指标，回答「算得准不准」。

临床评价。 产品在实际临床使用中的有效性和安全性，回答「用了有没有用、安不安全」。

技术指标好不等于临床有用。 比如灵敏度很高但特异度低的模型，可能产生大量假阳性，反而增加临床负担。

所以性能评价的结果要结合使用场景来解读 ——在什么人群中使用、阳性结果怎么处理、假阳性和假阴性各自的代价是什么。

建议在评价方案中就说明预期的使用场景，指标的选择和阈值的设定都应当与场景对应。

与人的配合

多数这类产品是辅助人做判断，而不是替代人。这带来额外的评价维度：

人机结合的表现。 人加模型的组合表现，与人单独、模型单独分别比较。

对人的影响。 模型的提示是否会影响人的判断，包括正面和负面的影响。

过度依赖的风险。 使用者是否会因为信任模型而放松自己的判断。

这部分的评价比单纯的技术指标更贴近实际效果。

我们的做法

这类产品的性能评价涉及数据与算法，与我们承担的物理与电气检测不是同一类工作。但含人工智能功能的产品仍然是器械，它的硬件部分、电气安全、电磁兼容、以及软件层面的通用要求，检测逻辑与其他器械一致。

我们承接这类产品时，关注的是软件版本与算法版本的记录、EMC 测试时的工作状态设定、以及基本性能的定义是否包含了算法输出的可靠性。 这几项是硬件检测与算法评价的衔接点。

如果你的产品含人工智能功能，需要安排硬件与电气相关的检测，可以把产品资料发过来一起讨论范围，或者直接联系：132 4819 8029。能力范围见服务介绍，送检要求见送检要求，流程见检测流程。

English version

Conclusion. Performance evaluation of devices based on machine learning differs in difficulty from conventional devices. Conventional device performance can be measured directly as a physical quantity: measure ten times and results should agree, largely independently of who measures and where. A machine learning model is different, its performance depending on the distribution of input data, so a model performing well on one dataset may perform markedly worse on another. The central question in evaluation is therefore what data to evaluate with and whether those data represent what will be encountered in use. With the wrong data, however good the reported metrics, they say nothing about real performance.

Distinctive difficulties. Data dependence, since performance correlates strongly with the test data distribution. Generalisation is hard to confirm, since behaviour outside the training distribution cannot be fully anticipated. Interpretability is limited, since the reason for a particular output is often hard to explain. Performance may change over time, declining as the data distribution shifts in use through changes in population, equipment or operating practice. Failure modes are not intuitive, since conventional device failures have definite physical causes while an erroneous model output often has no traceable one. And edge cases are hard to enumerate, since rare but important situations may be sparsely represented in the data. Together these mean that evaluation cannot rest on a single accuracy figure and needs a fuller body of evidence.

Dataset requirements. Independence requires complete separation from training data. Representativeness requires reflecting the actual population and setting. Diversity requires covering different sources, equipment and conditions. Scale requires enough data to support statistical conclusions. Annotation quality requires a reliable reference standard. And traceability requires clear provenance and processing. Independence is the most basic requirement and the one most often compromised: where test data overlap with training data, even as different samples from the same individual, measured performance comes out noticeably high and does not reflect real capability. Representativeness must be described specifically: how many sources, what population distribution, which equipment models, what acquisition conditions. Stating only how many samples a dataset contains leaves representativeness unjudgeable.

Establishing the reference standard. Evaluation needs a correct answer to compare against, and how that is established affects the conclusion directly. Consider the source of the reference standard, whether expert reading, another diagnostic method, follow-up or pathology. Consider the reliability of the reference standard itself, since expert readings also disagree and err. Consider handling of multiple readers, and how the reference is determined when readings differ. Consider reader qualification and blinding, meaning whether readers knew the model's output. And consider the reference standard's limitations, since an imperfect reference affects how metrics should be read. Blinding is critical to evaluation quality: where readers see the model's output before judging, the reference standard is contaminated.

Choosing metrics. A single metric rarely describes performance adequately. Accuracy is intuitive but misleading where classes are imbalanced. Sensitivity and specificity separate the ability to identify positives and negatives. Positive and negative predictive values relate to actual prevalence and sit closer to clinical interpretation. Receiver operating characteristic measures reflect overall behaviour across thresholds. Confidence intervals convey the uncertainty of the estimate. And stratified analysis shows behaviour across populations, equipment and conditions. Stratified analysis is a particularly valuable part of evaluation: good overall performance with markedly worse performance in a subgroup, such as an age band or data from a particular device, matters greatly for safety, and aggregate metrics conceal it.

Examining generalisation. External validation uses data from sources entirely different from development. Multi-centre data cover different institutions and equipment. Temporal externality uses data from different time periods. Edge case behaviour covers rare situations, poor image quality and atypical presentations. And adversarial input covers behaviour on abnormal input, particularly whether clearly wrong conclusions are returned with high confidence. External validation tells more about generalisation than internal testing, because internal test data, though independent, come from the same acquisition environment and population.

Monitoring in service. Because performance can change, post-market monitoring is necessary. Monitor the input data distribution for divergence from the development distribution. Assess performance continuously with new data at intervals. Collect user feedback, gathering and analysing error cases. And manage updates, verifying and notifying when the model changes. Distribution shift is a risk specific to these products: equipment updates, changed operating procedures and changing population composition can all move actual data away from the development distribution, with performance declining gradually and imperceptibly.

Organising the documentation. Performance evaluation documentation should include a detailed dataset description covering source, composition and distribution; the measures ensuring independence; the method for establishing the reference standard; the evaluation protocol with pre-specified metrics and criteria; overall results with confidence intervals; stratified analysis; analysis of edge cases; and the post-market monitoring plan. Pre-specified metrics and criteria must be fixed and recorded before evaluation: selecting favourable metrics to report after the fact is a methodological failing.

How we handle it. Performance evaluation of these products concerns data and algorithms and is not the same work as the physical and electrical testing we undertake. A product with artificial intelligence functions remains a device, however, and its hardware, electrical safety, electromagnetic compatibility and general software-level requirements follow the same testing logic as any other. For such products we attend to recording software and algorithm versions, setting the operating state for electromagnetic testing, and whether the definition of essential performance encompasses the reliability of algorithmic output. Those are the points where hardware testing meets algorithm evaluation.

Send us the product information and we will discuss the scope of hardware and electrical testing. Phone or WeChat: +86 132 4819 8029.