脑科学兴趣怎样变成 ISEF 课题：一次脑电选题拆解From an interest in neuroscience to an ISEF question: an EEG case study
脑科学的吸引力很直接：我们想知道，大脑在做什么，电信号又能告诉我们多少。但到了选题时，“用脑电做一个人工智能项目”仍然太宽。先买设备，还是先找数据？做疾病识别，还是做脑机接口？怎样判断一个想法已经有人做过？Neuroscience raises compelling questions about what the brain does and what electrical recordings can reveal. But “an AI project using EEG” is still a broad starting point. Should you buy a headset or find a dataset first? Study disease detection or a brain–computer interface? How do you tell whether an idea has already been explored?
这篇匿名产品案例记录 FairLab Pro 的一次选题分析。我们核对了十个 ISEF 往届项目，整理六个候选方向，再把其中三个写成可以比较、也可能被实验否定的研究问题。目前完成的是案例分析和研究设计，尚未开展这些候选课题的实验。This anonymized product case documents a topic analysis by FairLab Pro. We examined ten past ISEF projects, developed six candidate directions and expanded three into questions with testable comparisons. The work completed so far is case analysis and research design; the proposed experiments have not been run.
先看别人究竟验证了什么Start with what past projects actually tested
检索范围覆盖 2017—2026 年的脑电分析与脑机接口项目，包括运动想象、辅助沟通、癫痫检测和神经反馈。它是一组与选题相关的案例样本，不是对 ISEF 全部项目的统计。阅读时，我们更关心输入信号、实验对照、数据划分，以及摘要能支持多大的结论。The sample spans EEG analysis and brain–computer interface projects from 2017 to 2026, including motor imagery, assistive communication, seizure detection and neurofeedback. It is a relevant case sample, not a census of ISEF projects. We focused on the signals used, comparison conditions, data splits and the scope of claims supported by each abstract.
例如，2019 年的 BMCI-Net 项目 把 EEG 与肌电 EMG 结合，比较单一信号和融合信号的表现。这里值得借鉴的是对照设计：增加一种输入，是否真的比只用原来的信号更有效？如果只展示一个机械手能动，读者很难判断算法贡献在哪里。For example, the 2019 BMCI-Net project combined EEG with electromyography (EMG) and compared individual signals with their fusion. The useful lesson is the comparison: does adding another input improve performance over the original signal alone? A moving robotic hand, by itself, leaves the algorithmic contribution unclear.
2022 年的 P300 家居控制项目 使用线性判别分析（LDA），获得类别二等奖。它提醒我们，方法可以从简单模型开始。复杂网络是否必要，要靠实验回答。摘要中的高准确率仍需结合样本、任务和测试条件理解，不能直接当成真实家庭环境中的可靠性。The 2022 P300 home automation project used linear discriminant analysis (LDA) and received a second-place category award. A simple model can be a sensible starting point; experiments should establish whether a more complex network is needed. High accuracy reported in an abstract still depends on the sample, task and test conditions. It does not establish reliability in everyday homes.
2024 年的 运动想象控制项目 用脑电和卷积网络做实时游戏演示。这说明原型展示可以很直观，也让下一步问题更具体：第二天重新戴上设备，表现还一样吗？遇到眨眼、接触不良或陌生使用者时，系统会怎样？这些是我们据案例提出的追问，公开摘要没有给出完整答案。A 2024 motor imagery project used EEG and a convolutional network in a real-time game demonstration. That makes the prototype easy to understand and suggests further questions: does performance hold up when the headset is put on again the next day? What happens with blinking, poor contact or an unfamiliar user? These are our follow-up questions; the public abstract does not fully answer them.
另一个容易忽略的区别出现在 2025 年的 新生儿癫痫脑电项目：识别正在发生的发作，与提前预测未来发作，是不同任务。选题报告里多写一个“预测”，就可能改变标签定义、验证方式和整个研究难度。我们在整理案例时，把这种措辞差异也列入检查。The 2025 neonatal seizure EEG project illustrates another distinction: detecting an ongoing seizure differs from predicting a future one. Adding “prediction” to a proposal can change the label definition, evaluation and difficulty of the entire study. We checked wording as part of the case review.
筛到最后，留下三个值得先试的问题Three questions worth an initial experiment
我们的优先建议是从公开数据开始。这样可以先判断问题和方法是否成立，再决定是否需要设备或招募参与者。以下三条路线各有侧重；它们是待验证的选题建议，创新性还需要进一步检索和实验支持。Our recommendation was to start with public data, so the question and method can be tested before deciding whether equipment or participant recruitment is needed. The three routes below serve different interests. They remain proposals: their novelty needs further literature review and experimental support.
一、脑机接口什么时候应该暂缓输出？1. When should a brain–computer interface withhold an output?
假设系统正在识别左手或右手的运动想象。它每次都给答案，演示会很流畅；但信号很差时，强行给出的答案可能变成错误指令。一个可研究的问题是：同时考虑信号质量和模型不确定性，能否在保留相同输出比例的条件下，减少跨天测试中的错误？Consider a system that classifies imagined left- or right-hand movement. Answering on every trial makes for a smooth demonstration, but a forced answer from a poor signal may become an incorrect command. Can a combination of signal quality and model uncertainty reduce errors across recording days while retaining the same proportion of outputs?
实验至少要比较始终输出、只看模型置信度、只看信号质量，以及联合判断几种方法。关键是把输出比例对齐：拒绝大多数输入，当然容易提高剩余答案的准确率。我们希望知道，在同样愿意回答的情况下，哪种方法更可靠。The experiment should compare always answering, using model confidence alone, using signal quality alone and combining the two. Match the proportion of trials receiving an answer. Rejecting most inputs can make the remaining predictions look accurate; the useful question is which method is more reliable at the same output rate.
“让模型拒绝回答”已有研究，例如 2019 年的脑机接口保形预测论文。因此，创新点不能只写拒识机制，而应落在跨天变化下，质量信息是否提供了置信度之外的帮助。如果数据只有提示任务，也不能把结果直接写成连续使用中的误触发率。Abstention already has a research history, including a 2019 paper on conformal prediction for brain–computer interfaces. The potential contribution is whether signal quality adds information beyond confidence under changes between days. A dataset containing only cued trials also cannot establish a false-trigger rate during continuous use.
二、模型识别的是脑信号，还是眼动留下的线索？2. Is the model using brain activity or clues from eye movements?
这条路线更接近脑科学中的测量问题。头皮电极也会记录眼动、眨眼等活动。如果某种眼动刚好与任务标签一起出现，分类器可能学到一条捷径。准确率很好看，换一天或换个人却失效。This route focuses on measurement. Scalp electrodes also pick up activity from eye movements and blinks. If an eye movement happens to correlate with a task label, a classifier may learn a shortcut. It can score well in one setting and fail on another day or person.
可以设计一个受控实验：保持原始任务片段不变，添加已知的眼电成分，分别让眼电与标签保持相关、随机对应或反向相关，观察模型表现怎样变化。再用真实跨天数据检查：这种对捷径的敏感程度，是否能解释泛化能力下降？One controlled experiment would keep the original task segments fixed and add known ocular components whose relationship with the label is preserved, randomized or reversed. Then test on real recordings from different days: does sensitivity to that shortcut help explain a loss of generalization?
对照可以包括只用眼电的模型、原始脑电模型、清理伪迹后的模型，以及只使用运动相关区域通道的模型。已有伪迹研究 需要先读；清理后仍能分类，也不足以证明模型完全依赖神经活动。这项课题的价值在于把“可能有伪迹”变成一个可操纵、可检查的问题。Comparisons could include an EOG-only model, raw EEG, artifact-cleaned EEG and channels over motor-related regions. Prior artifact research is essential reading. Successful classification after cleaning is not proof of an exclusively neural source. The value of this question is making a suspected artifact measurable and experimentally controllable.
三、系统什么时候该重新校准？3. When should the system recalibrate?
脑电分布改变时，模型更新可能有帮助；如果变化只是短暂噪声，更新也可能把模型带偏。这里可以比较固定模型、定期校准、检测变化后校准，以及结合信号质量决定是否校准，在相同标注预算下看谁更稳定。Updating a model may help when the EEG distribution changes. If the change is temporary noise, updating may make it worse. Compare a fixed model, periodic recalibration, recalibration triggered by detected change and a trigger that also considers signal quality, all under the same labeling budget.
这条路线对实现要求更高。离线回放也必须按时间顺序，先预测，再更新，不能提前使用后续数据的统计量。欧氏对齐等已有方法 可以作为基线；新贡献需要落在更新时机及其代价上。This route is more demanding to implement. Even an offline replay must respect time: predict first, then update, without using statistics from future data. Existing methods such as Euclidean alignment can serve as baselines. The proposed contribution concerns when to update and the cost of doing so.
另外三个方向，为什么暂时放后面Why three other directions stayed on the shortlist
我们还保留了少通道条件下的电极故障容错、睡眠分期边界附近的不确定性，以及反馈延迟对注意任务表现的影响。前两者适合继续用公开数据探索，但要分别避免把“删掉记录通道”当成真实廉价设备测试，把睡眠阶段变化解释成意识机制。第三个方向可能需要参与者实验，应先核对 适用赛季的人类参与者规则，在所需审批完成后再招募或采集。Three other ideas remained on the shortlist: resilience to electrode failure with fewer channels, uncertainty near sleep-stage boundaries and the effect of feedback delay on attention-task performance. The first two suit public-data exploration, with limits: removing recorded channels does not test a real low-cost headset, and sleep-stage changes do not establish a mechanism of consciousness. The third may require participants. Check the applicable human participant rules and obtain required approval before recruitment or collection.
把第一个月写到可以开工Make the first month specific enough to begin
如果偏工程，我们会优先试“什么时候暂缓输出”；如果更想理解脑电测量本身，则先试眼动捷径。跨天校准适合在基础流程跑通后推进。这个排序依据的是问题能否清楚验证，以及实现成本，并不代表获奖概率。For an engineering emphasis, we would start with withholding unreliable outputs. For an interest in EEG measurement itself, we would start with ocular shortcuts. Recalibration is a sensible next step once the basic pipeline works. This ordering reflects testability and implementation effort, not award odds.
第一周，选定一个数据集，核对参与者、日期、通道和标签；第二周，跑通简单基线，按参与者或记录日期划分训练与测试；第三周，只加入一个要验证的改动，并补齐对照；第四周，查看每位参与者的结果和失败样本，决定继续、缩小问题还是换方向。数据预处理和阈值选择也要遵守同样的训练测试边界。In week one, choose a dataset and check participants, recording days, channels and labels. In week two, run a simple baseline with training and testing separated by participant or recording day. In week three, introduce one proposed change and its controls. In week four, inspect results per participant and failure cases, then decide whether to continue, narrow the question or change direction. Preprocessing and threshold selection must respect the same training–test boundary.
例如，BCI Competition IV 2a 数据集 包含 9 位参与者、两次不同日期的记录，可用于初步跨天实验；Lee 2019 运动想象数据 可作为后续扩展。选定前仍需检查许可、任务定义与实验问题是否匹配。公开数据降低了起步成本，严谨的数据划分仍然需要自己完成。For example, BCI Competition IV 2a includes nine participants with recordings on two different days, supporting an initial cross-day experiment. Lee 2019 motor imagery data offers a possible extension. Check licensing and task definitions before committing. Public data lowers the cost of starting; it still requires a carefully designed evaluation.
这次产品案例交付了什么What this product case delivered
这次分析留下了一份十个案例的对照报告、六个候选方向，以及三个较完整的实验设计。每个建议都带着需要验证的条件：最近的已有工作是什么，最简单的对照是什么，怎样的结果会让我们放弃原来的判断。The analysis produced a comparison of ten cases, six candidate directions and three more detailed experimental designs. Each recommendation carries questions to resolve: what is the closest prior work, what is the simplest meaningful comparison, and what result would make us abandon the original idea?
对 FairLab Pro 来说，这也是选题阶段值得交付的东西。目前，我们更愿意先验证跨天误判和眼动捷径：这两个问题能用已有数据起步，失败了也能知道假设错在哪里。是否值得继续，要看第一轮结果。For FairLab Pro, that is a useful deliverable at the topic-selection stage. For now, we would prioritize errors across recording days and ocular shortcuts. Both can be explored with existing data, and a failed experiment can help identify which assumption was wrong. The first results should determine whether to continue.
了解 ISEFLab Pro 的科研工作流 →Explore the ISEFLab Pro research workflow →
编辑说明：本文经匿名化重写，仅公开研究方法与选题判断；未公开个人资料、私人沟通或原始内部报告。项目事实链接至公开来源，候选课题不代表已验证成果。Editorial note: this case has been anonymized and rewritten to share research methods and topic decisions. Personal information, private communications and the original internal report are not published. Project facts link to public sources; proposed topics are not validated findings.