摘要：A tiny LLM predicts health events, even death, with remarkable accuracy.AI is learning the "grammar of disease," turning diagnoses

Artificial Intelligence and the Grammar of Life

人工智能与生命语法

How a tiny LLM tells an amazingly accurate clinical story.

一个小小的法学硕士学位如何讲述一个令人惊讶的准确的临床故事。

Updated September 18, 2025 | Reviewed by Michelle Quirk

更新于2025年9月18日 | 评论者：Michelle Quirk

A tiny LLM predicts health events, even death, with remarkable accuracy.AI is learning the "grammar of disease," turning diagnoses into predictive stories.There’s power in foresight, but also a kind of burden.

一个小小的法学硕士学位就能以惊人的准确度预测健康事件，甚至死亡。人工智能正在学习“疾病语法”，将诊断转化为预测故事。预见性是一种力量，但也是一种负担。

I came across a new study this week that really got me thinking. And the paper shined a new light on the value of large language models (LLMs) in medicine and even beyond. Researchers took a stripped-down version of GPT—a model with only about two million parameters—and trained it on individual medical diagnoses like hypertension and diabetes. Each ICD-10 code became a token, like a word in the sentence of a prompt, and each person’s medical history became a story unfolding over time.

本周我偶然发现了一项新研究，它深深地引发了我的思考。这篇论文为大型语言模型 (LLM) 在医学乃至更广阔领域中的价值提供了新的视角。研究人员采用了 GPT 的精简版——一个只有大约 200 万个参数的模型——并对其进行了高血压和糖尿病等个体医学诊断的训练。每个ICD -10代码都变成了一个标记，就像提示句中的一个单词一样，每个人的病史都变成了一个随着时间推移而展开的故事。

For a little context, GPT-4 and GPT-5 are believed to have hundreds of billions to trillions of parameters, making them hundreds of thousands of times larger than this small model. And yet, this tiny system was able to predict the “next word” in a person’s health story, including the next diagnosis, the next complication, and, with uncanny precision, even the timing of death. For me, that was a full stop. Let's dig in.

简单介绍一下背景：GPT-4 和 GPT-5 被认为拥有数千亿到数万亿个参数，这使得它们比这个小模型大数十万倍。然而，这个微型系统却能够预测一个人健康状况的“下一个词”，包括下一个诊断、下一个并发症，甚至以惊人的精度预测死亡时间。对我来说，这已经是句号了。让我们深入探讨。

Seeing the Arc of Illness

观察疾病的弧线

It's interesting to note that clinicians already think this way. A 50-year-old with hypertension might not alarm a doctor, but add diabetes and chronic kidney disease, and the physician starts to see the arc of possible futures that may include heart failure, dialysis, and even premature death. What this new model does is formalize and scale that intuition. It has seen hundreds of thousands of similar “patients” and knows, statistically, how their stories usually unfold and when.

有趣的是，临床医生已经这样思考了。一个50岁的高血压患者可能不会让医生感到担忧，但如果再加上糖尿病和慢性肾病，医生就会开始预见到未来可能发生的状况，包括心力衰竭、透析，甚至过早死亡。这个新模型的作用是将这种直觉形式化并规模化。它已经见过数十万名类似的“患者”，并且从统计学上了解他们的故事通常会如何发展以及何时发生。

This isn’t just predicting one outcome; it’s simulating a trajectory and forecasting which complications are likely to develop and how quickly they might appear. Think of it as if you were taking a snapshot of your current health and running it forward in time to see your clinical future.

这不仅仅是预测一种结果；它模拟了一种轨迹，预测哪些并发症可能出现以及它们出现的速度。想象一下，你正在拍摄一张你当前健康状况的快照，然后将其向前推进，以预测你未来的临床状况。

Predicting the Day You Die

预测你的死亡日期

One result was especially striking to me. The model could correctly distinguish who would die and who would live 97 percent of the time. That’s astonishingly accurate, especially given that the model was working with nothing more than a handful of diagnoses, age, and basic lifestyle factors.

有一个结果尤其令我印象深刻。该模型能够以 97% 的准确率区分哪些人会死亡，哪些人会活下来。这简直是惊人的准确率，尤其是在该模型仅基于少数诊断、年龄和基本生活方式等因素的情况下。

The Grammar of Disease

疾病的语法

Now let's talk about the power of diagnostic "language" in the context of a large "language" model. It seems as if AI is learning what might be called the grammar of disease. Each diagnosis is a “word,” each medical record is a “sentence,” and our lives are written in sequences of these tokens. Add hypertension and hyperlipidemia, and the model can already sketch a likely next chapter. Add diabetes or kidney disease, and the story becomes sharper, the ending more predictable. Each "word" adds more context and increases the statistical probability.

现在，让我们在大型“语言”模型的背景下，探讨诊断“语言”的力量。人工智能似乎正在学习所谓的疾病语法。每个诊断都是一个“词”，每份病历记录都是一个“句子”，我们的生活由这些词条的序列构成。加上高血压和高脂血症，模型已经可以勾勒出可能的下一章。加上糖尿病或肾病，故事就变得更加清晰，结局也更加可预测。每个“词”都增加了更多背景信息，提高了统计概率。

When AI Reads More Than Our Health

当人工智能读取的数据超出我们的健康范围时

So, let's push on this a bit. What does it mean, psychologically, ethically, and even practically, when a machine can read the story of our health and tell us what comes next or even how it might end? Do we really want to know our likely “final chapter,” even if we can’t change it? Or could this knowledge be used to rewrite the story and to intervene earlier, shift the trajectory, and add new chapters we never expected?

那么，让我们进一步探讨这个问题。当机器能够解读我们的健康状况，并告诉我们接下来会发生什么，甚至可能如何结束，这在心理、伦理甚至实践层面意味着什么？即使我们无法改变，我们真的想知道我们可能的“最终篇章”吗？或者，这些知识是否可以用来改写故事，更早地进行干预，改变轨迹，并开启我们从未预料到的新篇章？

Interestingly, this doesn’t stop at medicine. If a model can learn the grammar of disease, what about the grammar of psychology that looks at the sequence of experiences—captured in a diagnosis or word—that lead to depression or burnout? What about the grammar of relationships and the patterns that predict divorce or reconciliation? LLMs, trained on millions of human stories, already contain traces of these patterns.

有趣的是，这不仅仅局限于医学。如果一个模型可以学习疾病的语法，那么心理学的语法又如何呢？心理学研究的是导致抑郁或倦怠的一系列经历——这些经历体现在诊断或词语中。那么人际关系的语法以及预测离婚或和解的模式又如何呢？经过数百万个人类故事训练的法学硕士（LLM）已经包含了这些模式的痕迹。

So, this is where it gets both thrilling and unnerving for me. These models don’t just predict words; they predict the future shape of human lives, at least in a probabilistic or statistical sense. They hint at the possibility that every aspect of our existence, including our health, our choices, even our heartbreaks and heartbeats, has a statistical signature that can be read and projected into the future.

所以，对我来说，这既令人兴奋又令人不安。这些模型不仅仅是预测词语；它们至少在概率或统计意义上预测了人类生活的未来形态。它们暗示着这样一种可能性：我们生存的方方面面，包括我们的健康、我们的选择，甚至我们的心碎和心跳，都可能带有统计特征，可以被解读并投射到未来。

Which leaves us with a very human question. How much of this do we actually want to know? There’s power in foresight, but also a kind of burden. Perhaps the challenge isn’t just building models that can read our stories but in deciding when and how we want them to tell us what comes next.

这给我们留下了一个非常人性化的问题：我们究竟想知道多少？预见的力量在于此，但也是一种负担。或许，挑战不仅在于构建能够解读我们故事的模型，还在于决定何时以及如何让它们告诉我们接下来会发生什么。

来源：左右图史