内容来源：https://www.sciencenews.org/article/ai-health-side-effects-social-media

内容总结：

人工智能正尝试通过社交媒体捕捉药品不良反应信号。近日，一项发表于《PLOS数字健康》的研究显示，名为Waldo的人工智能系统通过扫描社交平台Reddit上43万条与大麻使用相关的帖子，成功识别出2.8万余条潜在不良反应报告。经研究人员核实，其中86%的帖子确实描述了真实存在的健康问题。

研究团队指出，公共卫生机构通常通过传统渠道收集药品不良反应，而社交媒体恰是官方监测的盲区。宾州理工学院信息技术专家理查德·洛莫特伊表示，这类平台能真实反映民众自发陈述的健康状况。加州大学圣地亚哥分校公共卫生研究员约翰·艾尔斯补充道，许多民众不熟悉官方报告渠道，更倾向在网络分享经历，"我们必须主动深入用户聚集的数字化空间"。

为训练该人工智能系统，研究团队首先采用1万条经人工标注的帖子作为训练样本。测试表明，Waldo在识别准确性上显著优于通用聊天机器人ChatGPT，后者误报率高出18倍，不过仍未能超越人工审核的精准度。

尽管该技术展现出应用潜力，研究者仍持审慎态度。洛莫特伊强调，人工智能在特定任务中表现优异，但未必能直接迁移至其他药品监测领域。他同时提出伦理考量，建议在严格隐私保护前提下，将AI识别的不良反应信息移交卫生部门，这对缺乏完善药物监管体系的国家尤为重要。

正如研究者所言，对遭遇罕见药物反应的个体而言，"即便发生率仅百万分之一，当发生在你身上时，就是百分之百"。这项技术未来或将成为连接公众与公共卫生服务的重要桥梁。

中文翻译：

人工智能能否识别社交媒体上的有害健康副作用？
这款人工智能工具通过扫描Reddit历史帖子来检测大麻引发的不良反应
“求帮助……我必须趴在地上惊恐发作20分钟才能平静……该去看医生吗？”这段求助出现在社交媒体Reddit上。发帖者吸食大麻后已持续数日出现恐慌发作症状。这类帖子通常不会引起公共卫生工作者的注意，但近期一项实验中的AI工具却捕捉到了这个信息。

这款名为“沃尔多”的工具扫描了43万余条Reddit论坛中与大麻使用相关的历史帖子，将上述求助帖与另外2.8万余条内容标记为潜在描述意外或有害副作用的帖子。研究团队对其中250条标记内容进行核查，确认86%的帖子确实反映了大麻产品引发的不良体验，该项研究结果已于9月30日发表于《PLOS数字健康》。若此类扫描技术得以普及，将有助于公共卫生工作者保护消费者远离有害产品。

宾州立大学信息技术专家理查德·洛莫蒂指出，这项研究的精妙之处在于揭示了研究人员能从美国疾控中心等政府机构尚未关注的渠道获取信息。尽管疾控中心等机构会开展问卷调查或收集自我报告的不良反应，但并未对社交媒体进行监测。而这里正是“人们自由表达的空间”。

参与开发“沃尔多”的加州大学圣地亚哥分校公共卫生研究员约翰·艾尔斯表示，许多人既无法及时就医，也不了解产品不良反应的官方上报渠道。当大量网民在线分享健康经历时，“我们必须主动深入他们的活动场域”。

密歇根大学医学院医学生卡兰·德赛解释，团队选择聚焦大麻产品是因为其广受欢迎却缺乏监管。“像我这样的20多岁年轻人，在中学和大学阶段就接触过JUUL电子烟、雾化器和大麻产品。了解使用者正在经历哪些副作用至关重要。”

为训练“沃尔多”，团队首先选取了1万条与大麻使用相关的Reddit帖子作为初始数据集，这些帖子已由其他研究人员手动标注出存在问题的副作用。德赛与同事用部分帖子训练系统后，在剩余帖子上进行测试。在此项任务中，该工具表现优于ChatGPT——这款通用机器人的误报率高出18倍，经常将未描述副作用的帖子错误标记。但相较于人工审核，AI仍存在差距。

所有这些准备工作都发生在主体实验之前，最终“沃尔多”成功标记了那条恐慌发作帖及数万条类似内容。

洛莫蒂指出，该工具在搜索其他药物、维生素或产品相关问题时是否同样有效尚待验证。针对特定任务训练的AI工具即使在高度相似任务中也可能表现不佳，对此他提醒“必须保持谨慎”。

尽管如此，洛莫蒂展望未来这类工具将协助监测社交媒体，但强调必须“以合乎伦理的方式”审慎推进。当用户发布罕见副作用信息时，系统可在隐私保护前提下标记问题并转交卫生部门。他认为这对缺乏完善药物副作用监测报告体系的国家尤为重要。

艾尔斯表示，终有一日，“沃尔多”这样的工具将帮助需要援助的人们与公共卫生工作者建立联系：“即便副作用发生率极低，一旦发生在你身上，就是百分之百。”

英文来源：

Can AI spot harmful health side effects on social media?
This AI tool scanned past Reddit posts to detect ill effects from cannabis
“Help me please … I can’t calm down without laying on the ground and freaking out for a good 20 minutes … Should I get medical help?”
This plea came from a post on the social media site Reddit. The person who posted the question had been having panic attacks for several days after smoking marijuana. Usually, this type of post goes unnoticed by people working in public health. But in a recent experiment, an AI tool was paying attention.
The tool, called Waldo, reviewed more than 430,000 past posts on Reddit forums related to cannabis use. It flagged the post above and over 28,000 others as potentially describing unexpected or harmful side effects. The researchers checked 250 of the posts that Waldo had flagged and verified that 86 percent of them indeed represented problematic experiences with cannabis products, researchers report September 30 in PLOS Digital Health. If this type of scanning became commonplace, the information could help public health workers protect consumers from harmful products.
The beauty of the work, says Richard Lomotey, is that it shows researchers can actually gain information from sources that government agencies, such as the U.S. Centers for Disease Control and Prevention, may not be looking at. The CDC and other agencies take surveys or collect self-reported side effects of illness but do not monitor social media. This is where “people express themselves freely,” says Lomotey, an information technology expert at Penn State.
Many people don’t have access to a doctor or don’t know about the official way to report a bad experience with a product, says John Ayers, a public health researcher at the University of California, San Diego in La Jolla who worked on Waldo. Lots of people share health experiences online. “We need to go where they are,” he says.
Karan Desai, a medical student at the University of Michigan Medical School in Ann Arbor, says the team chose to focus on cannabis products because they are very popular yet largely unregulated. “People in my age demographic, in their 20s, grew up in high school and college with these JUULs, these vapes, these cannabis products,” he says. “I think it’s important for us to know what side effects people are experiencing with using these.”
To prepare Waldo, the team began with a smaller group of 10,000 different Reddit posts about cannabis use. Other researchers had gone through these and identified problematic side effects by hand. Desai and colleagues trained Waldo on a portion of these posts, then tested it on the remaining ones. On this task, the tool outperformed ChatGPT. The general-purpose bot marked 18 times more false positives, indicating posts contained side effects when they didn’t. But it did not outperform the human reviewers.
This all happened before the team’s main experiment, in which Waldo tagged that panic attack post and tens of thousands more.
It remains to be seen whether Waldo would work as well searching for issues related to any kind of drug, vitamin or other product, Lomotey says. AI tools trained on one task may not work as well even on very similar tasks. “We have to be cautious,” he says.
Still, Lomotey imagines a future where tools like Waldo would help keep an eye on social media. This would need to be done carefully, “in an ethical way,” he says. When a person posts about a rare side effect, such tools could flag the issue and pass it on to health officials, with privacy protections in place. He imagines that this could be especially useful in countries that don’t have robust systems in place to monitor and report on drug side effects.
Someday, tools like Waldo might help link people who need help to the public health workers who can provide it. “Even when [side effects] can be rare, when they happen to you, it means all the world,” Ayers says.