保险理赔为何是 AI 最棘手的数据问题：Wisedocs 的 Denys Linkov

简要说明Brief Description

保险理赔是 AI 最棘手的数据问题——Wisedocs 的 Denys LinkovInsurance Claims Are AI's Messiest Data Problem — Denys Linkov， Wisedocs

目录Table of Contents

• 简介
  Intro
• 保险业的技术转型
  Insurance's technology transformation
• 索赔的实际运作方式
  How a claim actually works
• 第 1 步：清理数据
  Step 1: clean data
• 步骤 2：文档内分析
  Step 2: analysis within documents
• 在医疗记录上对模型进行基准测试
  Benchmarking models on medical records
• 第 3 步：跨主张的洞察
  Step 3: insights across claims
• 处理栈
  The processing stack
• 监管、成本与未来展望
  Regulation, cost and what's next

简介Intro

好，感谢大家来参加今天最后一场研讨会。我叫丹尼斯，今天想和大家聊聊保险理赔数据，以及它对企业级 AI 项目能发挥多大的作用。那我们开始吧。现场有谁在保险行业工作？有几位。那有谁想进入保险行业？好，有意思，也有几位。好吧，保险是我们经济中一个非常有意思的行业，人们往往是在机缘巧合下进入这个行业的，但它却支撑着我们生态系统中非常重要的一部分。All right, thank you for joining the last workshop of the day. So, I'm Dennis. I'll be sharing a little bit about insurance claims data and how powerful it can be for enterprise-wide AI programs. So, let's get started. So, who here works in insurance? Few people. Who here wants to work in insurance? Okay, interesting. Few people. All right. So, insurance is a very interesting sort of sector of our economy where people sort of find their way working there, but it powers such a important part of our ecosystem.

接下来，我们会简单谈谈保险的整体情况。最终，本次讨论的重点将是如何全面改善我们的生活。因此，我们会聚焦医疗理赔，并将改善理赔结果和患者预后作为本次讨论的重点，我也会不断回到这一主题。So, we're going to talk a little bit about insurance as a whole. And we're ultimately going to focus this talk on how do we improve our lives in general? So, we'll focus on medical claims, we'll improve the claim and patient outcomes as a focus of this talk, and I'll keep bringing that back.

我觉得，在很多金融科技类的金融保险讨论中，有时大家会过于关注商业层面。所以我总是希望大家把注意力放在这一点上。简单介绍一下我自己：我叫丹尼斯，在 Wisdocs 负责 AI 项目。我们帮助企业处理复杂的医疗理赔。我担任这个职位已有两年，从 2024 年以来，我们见证了巨大的变革。I think often times in a lot of fintech sort of finance insurance discussions, sometimes things can get a little bit too business focused. So, I always want everybody to focus on this aspect. So, a little bit about myself. I'm Dennis. I lead the AI program at Wisdocs. We help companies with their complex medical claim processing. Been in the role for 2 years now, and we've seen a huge amount of transformation since 2024.

所以，你可能已经看到了一些关于医疗保障和保险总体情况的新闻报道。GLP-1 带来的变革已经改变了我们的经济和医疗保健环境，也改变了保险定价的方式。So, you might have seen some of the news coverage in terms of medical coverage and insurance in general. The whole GLP-1 transformation in our economy, in our health care environment has changed how insurance is pricing things as well.

保险业的技术转型Insurance's technology transformation

我们还看到，配偶保险以及更广泛的保险保障一直在不断发展，成本上升等因素也在改变保险业务的运作方式。因此，这次演讲将围绕三个主要部分展开。第一部分是我们正在经历的技术转型。第二部分，我会稍微介绍一下保险领域的一些具体细节。We also see that spouse insurance and broadly insurance coverage has continued to evolve with increasing costs and other factors that change the approach of insurance. So, the talk is going to focus on three main chapters. The first one is on the technology transformation that we have. The second is I'll go a little bit into the insurance space in terms of some of the call it nitty-gritty, but sort of a brief introduction there.

然后，我还会谈谈 AI 将如何帮助改善保险服务和患者体验。我们来聊聊技术转型。我们可能都已经在各自的工作中，在公司内部启动了某种形式的数字化转型，无论我们是在小型初创公司工作，还是在大型企业工作。现在，面临的挑战是，我们从未完成过这一转型。And then also I'll talk about how AI is going to help make insurance and the patient experience even better. Let's talk about technology transformation. So, we all probably in our work have started some form of digital transformation within our companies, whether we work in smaller startups or with in larger enterprises. Now, the challenge with that is that we never finished this transformation.

AI 在 2023 年和 2024 年席卷而来，但大多数公司还没完成云迁移，还没充分发挥大数据的潜力，也还没有部署好一个干净的数据湖，对吧？在座各位，有谁拥有一个让自己引以为傲的数据湖？好吧，寥寥无几。真的有可能拥有一个让自己引以为傲的数据湖吗，对吧？除此之外，现在关于 AI 的炒作铺天盖地，我们必须确保先具备其他必要条件，才能真正部署实际应用。AI hit us in 2023 and 2024, but most companies have not finished their cloud migrations, have not finished leveraging big data to its full potential, have not deployed a clean data lake, right? Who here has a data lake they're proud of? All right. Very very few people. Is it even possible to have a data lake you're proud of, right? And on top of that, we're now getting all this hype about AI, and we need to really make sure that we have the other prerequisites down before we can actually deploy real applications.

所以，在保险业和许多其他行业，纸张曾是处理事务的主要方式，对吧？处理一项医疗理赔时，你会拿到一叠文件。你会把它们整理好，把它们放在一张张巨大的桌子上，还会用卡片把它们分到不同的部分。要从理赔本身获取任何其他信息，实际上非常困难，因为所有东西都是纸质的，要怎么进行跨理赔分析呢，对吧？So, in insurance and many other industries, paper was the big way of processing things, right? You'd get a stack of papers of going through a medical claim. You'd organize it. You'd have these giant tables full of them. You'd use cards to put in certain sections. And this is very difficult to actually get any anything outside of the claim itself, because how do you do cross-claim analysis when everything is paper, right?

你可以放任意多的文件夹，但这非常困难。如今，我们仍未彻底告别纸质时代。世界各地仍有不少场景在使用纸张，无论是在医生诊所、法务部门，还是保险公司内部。因此，我们还没有完全摆脱纸张，但已经越来越接近了。说到数字化，我们也还没有彻底告别传真机。传真市场仍在持续增长。You could put as many folders as you want, but it's very difficult. Now, we still have not exited the paper era. There are still a number of practices that they use paper around the world, whether it's in the doctor's office, in legal departments, or within insurance companies themselves. So, we're still not over paper, but we're getting close. Now, in terms of digitizing, we are still also not over the fax machine. The fax market is continuing to grow.

全球规模超过70亿美元，这总是令人惊讶。我去年在 An Engineer SF 的演讲中提到过这一点，情况至今没有变化。传真仍在使用。如今，我们正设法把传真接入数字系统，以便将其转换成相当规范的格式，但这仍然无法解决我们如何在医疗保健系统中真正交换信息的问题。It's over $7 billion globally, which is always surprising. I referenced this in my talk at An Engineer SF last year and it has not changed. Fax is still being used. Now, what we're doing now is we're jerry-rigging faxes into our digital systems where we can convert them in a fairly clean format, but it still does not solve the problem of how we actually exchange information in the health care system.

现在，如果我们想想技术被采用的速度，电子健康记录（EHR）在2008年的采用率只有8%。直到新冠疫情期间，EHR的采用率才真正达到80%和90%的水平。再说一次，许多医院仍在使用纸质记录。所以，我们还没有完全走过数字化、数字转型的时代，可人人都想做人工智能。如果我们想想人们在处理理赔时会收到什么类型的表格，下面是三个大州用于工伤赔偿的三份样表。Now, if we think about how quickly technology has been adopted, EHRs, electronic health records, uh we're only at 8% adoption in 2008. And it only with the COVID pandemic did we actually hit the '80s and '90s in EHR adoption. And again, many hospitals still run on paper. So, we're still not fully through this digitization, digital transformation era, and yet everybody wants to do AI. If we think about the types of forms people receive when when doing claims processing, these are some three sample forms from three large states for workers compensation.

所以，工伤赔偿是指你在工作中受伤后，提交一份申请，说：“嘿，我摔倒时伤到了胳膊，或者我当时在开叉车，结果出了点问题。”而且这些表格都不一样，对吧？美国有50个州，通常也就有50种不同类型的表格。So, workers compensation is if you get hurt at work, you file a claim saying, "Hey, I hurt my arm falling down or I was driving a forklift and then something went wrong." And none of these forms look the same, right? So, within 50 different states in the US, you have 50 different types of forms generally.

它们询问的信息各不相同，并没有统一标准。不过，它们在一定程度上还是标准化的，因为收到的每份表格看起来都一样，所以在这方面可以有一定的确定性；但即便如此，所包含的信息还是不同的。所以，即便是在美国这样一个国家内部，更不用说世界其他国家，我们收集信息的方式也不尽相同，而且这些版本有时也会发生变化。如果要说保险领域的人们关心什么，那就是赔付率，对吧？They all ask for different information. They're not standardized. They're fairly standardized in that every form that comes in looks the same, so you can have some determinism there, but even then the information is different. So, even as a country in the US, let alone any other country in the world, we don't capture information in the same way and these versions sometimes change as well. If we talk about what people care about in the insurance space is a loss ratio, right?

那么，你收取了多少保费，又赔付了多少钱？我会再多谈一点这些数字以及它们的含义，但总的来说，这些因素决定了一个组织的最终盈亏，对吧？在公布季度或年度业绩时，你会重点说明你的赔付率是多少、预测风险的能力有多强、承保做得有多好、欺诈管理得有多好，以及费用管理得有多好。So, how many premiums are you collecting and how much money are you paying out? So, I'll talk a little bit more about the these numbers here and what they mean, but in general, this is what drives the bottom line of an organization, right? When you report a quarterly earnings or an annual earnings, you're going to be emphasizing what is your loss ratio, how well can you predict risk, how well can you underwrite, how how well are you managing fraud, how well are you managing your expenses.

所以，公司都想降低成本、多赚钱，对吧？这就是一般意义上的资本主义。嗯，如果我们看看这里的一些统计数据，通常我们的赔付率大约在 50% 左右，也就是说，我们需要为可保索赔支付的金额，与我们收到的保费相比，大约是这个比例。现在，我介绍了这些背景信息，到目前为止，在我的演讲中只提到过一次 AI。And so companies want to keep costs down and make more money, right? That's just capitalism in general. Um and if we think about some of the statistics here, we generally uh hover around the 50% um loss ratio here in terms of the amount of money we have to pay out for insurable claims uh compared to the premiums we get. Now I'm I'm introducing all of these sort of pieces of context and I've only said AI once so far in my talk.

再强调一次，认真思考我们试图解决的问题，然后只有在 AI 确实是合适的解决方案时才选择它。要获得合适的 AI 解决方案，我们需要经历适当程度的转型，对吧？我们需要明确正确的目标。我们需要有一支团队，能够清楚地阐述我们想用 AI 做什么。不能只是让人们上台喊着“AI、AI、AI”，对吧？And it's really important again to think about the problem that we're trying to solve and then only choose AI if it's the right solution. Now for AI uh to get proper AI solutions, we need to go through the right amount of transformation, right? We need to have the right goals in mind. We need to have a team that can actually articulate what we want to do with AI. We can't just have people get on stage and say AI AI AI, right?

我们需要具备合适的人员技能，才能真正完成这项转型；同时也需要建立合适的框架，无论是治理、安全、基础设施管理还是数据方面的框架。现在有个问题。嗯，在座各位中，哪些人在企业环境下，只要提出申请，就能在一周内拿到一个数据库或虚拟机？好。哪些人做不到？好，这已经是进步了。在座各位中，哪些人能在一周内将应用部署到生产环境？We need the right people skills to actually accomplish this transformation and we need the right frameworks in place, whether it's governance, security, infrastructure management, and data. Now there's a question here. Um for those working at an enterprise context, who here could get a database or a VM within a week if they requested it? Okay. Who could not? Okay, that's progress. Who here could deploy an application to production in a week?

一个新应用。好吧。谁做不到呢？好吧。所以我们还是得有所进展。谁能在一周内选定一家新的 AI 供应商？啊，这就对了。没人举手。没错。我们仍然需要时间来完成所有这些不同的流程。嗯，在数字化转型方面，它们肯定比 10 年、20 年前好多了，但很多公司仍然受困于治理和安全问题。A new application. Okay. Who could not? Okay. So we still need some progress. Who can choose a new AI vendor in a week? Ah, there we go. No hands. Right. So it still takes time for us to go through all these different processes. Uh they're certainly better than they were 10, 20 years ago in the digital transformation component, but a lot of companies still get stuck in governance and security.

其中很多都与这里的这些因素有关。这里有多少人曾被高管要求做过类似的查询？“给我一份完美的数据集。”对吧？有几个人，对吧？我们以前也遇到过这种情况。“嘿，你能不能赶紧帮我把那份报告调出来？”嗯，现在有了 AI，这要容易得多，但你得先假设这些数据确实存在，对吧？那么，如果我们在这里提出一个更复杂的查询呢？比如说，我们有一项理赔，针对某位患者及其病情——在这个例子中，是脚踝受伤——我们想确保能把患者分配给一位患者治疗效果排名前50%的医生，对吧？And a lot of those have to do with the factors here. Who here has been asked for some kind of query like this from an executive? Give me some data set that is perfect data. Right? A few people, right? We've seen this before. Hey, can you pull that report for me quickly? Um it's much easier now with AI, but you need to assume that data exists, right? So, if we come up with a more complex query here, right? Uh let's say that we have a claim and for a given patient and their condition, uh in this case, uh an ankle injury, we want to make sure that we can assign them to a doctor who has top 50% tile uh patient outcomes, right?

在某个近期的时间范围内。现在，这个查询相当复杂，但如果你能直接在自己的数据集上编写这个查询并得到结果，那就可以了，对吧？Within some recent time frame. Now, this query is fairly complex, but if you could actually just write this query on your data set and get that result, right?

我会觉得很惊讶，对吧？在企业环境中，数据处理非常、非常具有挑战性。因此，要取得这样的成果并获得这类洞见，我们必须以非常明确的方式构建数据系统。我们来稍微谈谈医疗理赔。一般来说，我按照 NAIC 的分类对保险进行了归类。我们有财产与意外险、人寿险、意外与健康险，以及范围更广的健康保险。I'd be amazed, right? Data is very very challenging in an enterprise environment. So, for us to get these kind of outcomes and get these kinds of insights, we have to structure our data systems in a very well-defined way. So, let's talk a little bit about about medical claims. So, in general, I've grouped uh insurance based on the NAIC uh classifications. We have property and casualty, life, accident, and health, and sort of broader health uh insurance.

我实际上会讲财产与意外险，因为这类保险的理赔通常最复杂。说到财产与意外险，嗯，我主要从事的是意外险这一块。比如，呃，汽车保险、各种类型的责任险，包括一般责任险、产品责任险和工伤赔偿险。另外还有整个财产险领域，呃，它也有自己的一套复杂之处。I'll actually be talking about property and casualty, since there the claims are typically the most complex. If you think about P&C, um I've mainly worked on the casualty side. So, uh auto uh insurance, uh various types of liability, general liability, product liability, workers' compensation. There's also the whole property side, uh which is its own uh set of complexities.

不过，通常来说，这是一个万亿美元的市场，对吧？呃，按保费计算，而再保险则有其自身的一套形式。But, generally, this is a $1 trillion market, right? Uh based on premiums, and reinsurance is its own kind of forms.

索赔实际是如何运作的How a claim actually works

所以，如果我们想想这些蓝色方框，其中许多都处理复杂的医疗索赔。有人因为产品有缺陷而受伤，有人遭遇车祸，呃，还有人在工作中受伤。现在，审核索赔并向患者或索赔人作出答复，需要很长时间。最理想的情况下，也要几周；最糟糕的情况下，则可能拖上几年。这取决于情况有多复杂、监管环境有多复杂，以及诉讼有多频繁。So, if we think about these blue boxes, many of these handle complex medical claims. Somebody gets hurt because a product is defective, somebody gets into a car accident, uh somebody gets hurt at work. Now, adjudicating a claim and getting a response back to a patient or a claimant takes a long time. In the best case scenario, it takes a few weeks. In the worst case scenario, it can drag on for a few years. Depending on complex the scenario is, depending on complex the regulatory environment is, depending on how litigated the environment is.

而且，即使医疗服务已经提供，理赔也可能一拖再拖。所以，这并不是一个简单的流程。我希望我们都能想象一下，如果能把这一过程缩短十倍，医疗体系的效率会提高多少，对吧？And even once you get the medical services delivered, the claim can keep dragging on and on. So, this is not a simple process, and I want all of us to imagine how much more effective the healthcare system could be if we could shrink this by a factor of 10, right?

你在工作中受伤后，提交申请，嗯，走完那套流程，一切都是数字化的，而且进展顺畅。那么，为什么还是要花这么长时间呢？即使只看医疗这一块，也有很多环节需要交接，更不用说法律方面和实际的保险方面了，对吧？所以，当你在这类情况下受伤，并且需要为一项复杂的理赔申请保险赔付时，你得先拿到诊断结果，还得去看各种专科医生。You get hurt at work, you submit the application, uh you go through that process, and everything is digital, and everything is smooth. So, why does it take so long? There's a lot of handoff points with getting even the medical side, not accounting for the for legal side, the actual insurance side as well, right? So, when you when you get hurt uh in one of these scenarios and have to be insured for a complex claim, you have to get the diagnosis, you have to go to various specialists.

有时候专科医生会意见不一致，对吧？那你就得确保自己有合适的保障，嗯，不管是初级保健，还是这些额外的责任险种。还有医疗服务利用审核，在审核中，医生必须彼此认同某项治疗程序确有必要。嗯，我们有治疗，也有账单。Sometimes specialists disagree, right? Then you have to make sure that you have the right coverage, uh whether it's primary care or one of these uh additional forms of liability. You have utilization reviews where doctors have to agree with one another that a procedure is necessary. Uh we have treatments, we have billing.

你会试着弄明白账单是什么意思，对吧？还要看福利和保障说明。你会收到付款，也会继续接受治疗。呃，这里还有很多可以添加的额外环节，也有许多岔路，嗯，具体取决于各个步骤，但总的来说，这不是一个简单的流程。一路上，很多事情都可能出错。呃，我刚才随口就列出了大约15个问题，但这涉及人力资源方面的情况，也涉及保险公司理赔专员的高流动率和高退休率，还涉及技术问题——人们一遍又一遍地重新处理相同的数据。You try to figure out what your bill means, right? With the explanation of benefits and coverage. You get the payments, and you have continued treatments. Uh there's so many additional bubbles here that can be added and forks in the road, um depending on the steps, but in general, it is not a simple process. And along the way, lots of things can go wrong. Uh I just listed around 15 that come to mind immediately, but this has to do with the HR situations and insurance companies with claim adjusters having high turnover, high retirement rates, has to do with technology of people reprocessing the same data over and over again.

这跟我们最终会遇到的那类情况有关：索赔人压力很大，有时会犯错，处于无法理性行事的状态，对吧？Has to do with the types of situations we end up with where claimants are stressed and sometimes make mistakes and are in the position where they're not acting rationally, right?

因为，如果你经历过一场严重的车祸，你就不会是最理性的人。所以，我们来谈谈技术如何帮助我们。接下来我们看一个工伤赔偿申请。有人以前见过工伤赔偿申请吗？有几个人，好。通常有人受伤后，必须在工作场所内报告，并提交相关申请。我展示的那三份表格，分别是德克萨斯州、纽约州和加利福尼亚州的，都是典型的受理表格。Cuz if you're you've gone through a traumatic car accident, you're not going to be the most rational person. So, let's talk about how technology can help us. So, let's go through a workers' compensation claim. Has anybody seen a workers' comp claim before? Few people, okay. So, typically when somebody gets injured, they have to report it within the workplace and then make a filing of that as well. Those three forms I showed, the Texas, New York, and California ones, those are typical intake forms.

然后你会经历这样一个过程，实际去看各种各样的人，比如去急诊室，接受其他治疗，等等。而且你不一定会马上审核自己的理赔申请，可能要等到之后收到理赔申请时才会审核。经历这个过程时，会涉及各种类型的理赔申请。这里我生成了一些合成的理赔申请作为示例，不过它们和生产环境中的情况相当接近。And you go through this process of actually visiting various people, whether it's an emergency room, whether you have additional treatments, and so forth. And you don't always have your claim reviewed until you get claims later on. Now, as you go through that, there's various types of claims. Here I generated some synthetic ones as an example, but they're they're fairly close to what you have in production.

所以，你会经历各种类型的治疗、接受不同类型的医疗程序；你可能选择在网络内就医，也可能选择网络外就医，还可能为了这些治疗出行，有时还会遇到相互矛盾的信息。随着我们逐一经历这些流程，会发现所有文件的样式都不一样。每家医院都有自己的账单格式，对吧？每位医生写病程记录的方式也都不同。So, you go through all these types of treatments, you get different types of procedures, you might be in network, you might be out of network, you might be traveling for these, and sometimes you have conflicting information as well. And as we go through each of these processes, all the documents look different. Every hospital has its own way of structuring a bill, right? Every doctor has a different way to write a progress note.

它完全没有任何一致性。所以，数据是非结构化的，而且很多时候并不是通过电子数据流传递的，而是 PDF、电子邮件或传真。因此，除了底层的业务复杂性之外，还叠加了种种技术复杂性。然后你会遇到这样的情况，对吧？有人做了一次评估，却得出了三个不同的结果，对吧？It's not consistent in any way, shape, or form. So, this is why the data is unstructured and many times it's not in an electronic feed. It's a PDF, it's an email, it's a fax. So, you have all these technology complexities that layer on top of all the business complexity that lies underneath. Then you have something like this, right? Somebody does an assessment and there's three different results, right?

比如说，某人关节的活动范围是多少。你要如何解决这个问题，才能对索赔作出裁定？而且我们甚至还没把另一部分因素考虑进来。在这些医疗流程进行的同时，还存在诉讼风险。对于涉及诉讼的索赔，保险公司通常站在被告一方，而另一方则是原告。原告一方有动力尽可能多地处理案件，并争取尽可能高的和解金额。For let's say the range of motion that somebody has. How are you going to resolve this in order to adjudicate the claim? Now we haven't even layered the other portion on top of this. While these medical processes are going through, there's also the risk of litigation. So with an insurance, typically insurance is on the defendant side of a litigated claim and you have the plaintiff side. Now the plaintiff side is incentivized to handle as much volume as possible and to get as big of a settlement as possible.

所以你确实会陷入这样一种局面：存在两套不同的激励机制。原告一方会尽可能加快进度，尽可能提出金额最高的索赔，到处竖起巨大的广告牌。我们飞过来的时候大概就看到了，对吧？到处都是。So you actually get into a situation where you have two different incentive structures. So on the plaintiff side, go as quickly as possible, make the most the biggest claims possible, put up giant billboards everywhere. We saw those when we were flying in probably, right? all over the place.

而且，与保险方或被告方相比，原告方实际上更快地采用了 AI。因为你能拿到的报酬比例取决于办案速度以及索赔金额的大小。根据所在司法辖区的不同，律师实际获得的赔偿金比例也会有所不同：在加州，工伤赔偿案件的律师费比例可能为 9% 至 15%；在其他司法辖区，则可能达到三分之一，甚至 40%。And the plaintiff side has actually adopted AI a lot more quickly compared to the insurance or defendant side. Because you get a percentage of how fast you move and how big your claim is. Now the range of the actual percentage of compensation for lawyers depending on your jurisdiction can range from 9 to 15% if you're in California for workers comp fees. In other jurisdictions, you get into that 1/3 territory or even 40%.

所以现在，除了医疗保健、保险、技术方面的挑战和行政事务之外，你还要面对诉讼，而保险公司会为诉讼索赔预留大量资金。现在的挑战在于，如果你没有进行适当的数字化转型，就特别容易受到不法分子的侵害，对吧？他们可以反复向你提交索赔，也可以提交欺诈性索赔，而现在要做到这一点甚至更容易。So now on top of healthcare, on top of insurance, on top of technology challenges, on top of administration, you now have litigation and insurance companies settle aside a lot of money for litigated claims. Now the challenge here is that if you have not gone through a proper digital transformation, you are especially vulnerable to bad actors, right? So you can be sent repeated claims, you can be sent fraudulent claims as making that even easier to do right now.

所以，如果你没有能力真正处理这些问题并高效地完成流程，系统里的欺诈就会越来越多，对吧？这就像你的网站很容易遭到 DDoS 攻击一样，对吧？人们就会不断对你发起 DDoS 攻击。而当我们开始解决所有这些子问题时，就会忘记什么才是最重要的，对吧？所以，就某个人所接受的护理质量、降低成本、确保流程快捷顺畅以及提供良好体验而言，这些才是最重要的方面。So if you don't have the ability to actually handle those and process things efficiently, you get more and more fraud in the system, right? It's similar to if your website is easy to DDoS, right? People are going to keep DDoS-ing you. And we forget about what matters most as we start solving all these sub-problems, right? So, these are the things that matter most in terms of the quality of somebody's care, reducing costs, making sure it's a quick process, and it's a good experience.

步骤 1：清理数据Step 1: clean data

那么，我们来逐一看看并解决其中的一些问题。我会谈谈三个不同的阶段，以及它们能如何提供帮助。首先，我们来谈谈干净的数据，以及如何实现这一点。其次，我们来谈谈文档内分析，然后再谈谈跨理赔因素。好，先说干净的数据。这里有人以前处理过理赔数据或任何类型的保险数据吗？有几位。好，你们处理过或见过多少重复记录？So, let's go through and solve some of these problems. I'll talk about three different phases and how they can help. So, number one, let's talk about clean data, how we can accomplish that. Number two, let's talk about within document analysis, and then let's talk about cross-claim factors. Okay, clean data. So, has anybody here worked with claims data before or any kind of insurance data? Few folks. Okay. How many duplicate records sort of have you processed or have you seen?

很多？是啊，数百万，对吧？呃，这取决于你所涉及的司法辖区，比例可能高达 50%，甚至更高，对吧？所以会有大量重复项涌入，有时这是故意的，对吧？有时为了让理赔得到处理，你确实得不断催促别人，对吧？跟进我的理赔，跟进我的理赔，对吧？Many? Yeah, millions, right? Uh depending on the jurisdiction that you work with, they can be as much as 50% or even higher, right? So, you have so many duplicates coming in, and sometimes this is intentional, right? Sometimes to get a claims processed, you actually have to keep pinging people, right? Following up with my claim, following up with my claim, right?

嗯，所以因为人们处理得慢，就会出现我们在技术领域看到的这种连锁故障效应。病历会和其他患者的记录混在一起，对吧？比如有个鲍勃·乔伊，还有个小鲍勃·乔伊，对吧？这种情况相当常见，导致他们的病历混在一起。这违反了 HIPAA 规定，会带来一大堆监管风险，但现有系统和数字化转型并没有针对这种情况做好准备。而且，当别人的病历混进来时，实际上可能会据此作出决策。Uh so, because people are slow, you get into this sort of cascading failure effect that we see in technology. Commingled records with other patients, right? You have Bob Joe and you have Bob Joe Jr. Right? Those medical records get mixed in a at a fairly often occurrence. It violates HIPAA. There's a whole bunch of regulatory risk for that, but the system and the digital transformation is not set up for that. And decisions can actually be made when other people's medical records are mixed in.

对吧？所以，当理赔案件特别复杂时，这一点尤其敏感，因为很容易遗漏信息。上游的医生或管理员一旦把某人的病历混了进去，整个理赔案件就会突然出问题。然后还有一些更常见的情况，比如资料扫描质量差或尚未数字化。Right? So, this is especially sensitive when there's a high uh complexity claim where things can be missed. Somebody's record gets mixed in by a doctor upstream or administrator and suddenly the whole claim falls apart. Then we have the more expected items of poorly scanned or undigitized data.

对吧？有些手写字你都看不懂，更别说医生的字迹了，对吧？有些东西是以纸质形式送来的。所以，我们不断遇到这些质量不佳的数据示例，而它们出现的原因有很多，对吧？其中一些情况是因为医疗系统里有很多复制粘贴，对吧？医生或护士更新病历时，会经历这个过程，对吧？Right? You have some some handwriting that you can't understand, let alone doctor's handwriting, right? Some things come in on paper. So, we keep getting these poorly uh clean data examples and they happen for many reasons, right? So, some of this happens because uh you have a lot of copy pasting that happens in the medical system, right? You go through uh doctors or nurses as they're updating the medical records, right?

他们会在患者之间复制粘贴，也会在不同的就诊记录之间复制粘贴。呃，你们可能亲身经历过，但这在这个领域里很常见。呃，电子病历系统的设置其实并不统一。那么，在座各位有人见过电子病历吗？比如自己的病历，对吧？有人换过医疗服务提供者或医疗网络吗？嗯。它们看起来都一样吗？They'll copy paste between patients. They'll go through and they'll copy paste between visits. Uh you might have seen that personally, but it happens a lot uh in the space. Uh EMRs are actually not set up consistently. So, has anybody here seen an EMR before? Their own medical record, right? Anybody here moved across providers or networks? Yeah. Do they look the same?

不，它们看起来并不一样。嗯，我来自加拿大多伦多，我们这里很多医院每次部署电子病历系统（EMR）时，都会从头开始实施。这种情况往往既是因为医院内部原有的流程，也会在整个医疗网络中出现。为什么会以某种特定方式进行设置，背后有着非常有趣的动态。如果你想聊聊数据的整洁度，人们会在 EMR 的嵌套字段里做一些非常奇怪的事。No, they don't look the same. Um I'm from Toronto, Canada and many of our hospitals do fresh implementations every time they set up an EMR. And this often happens both due to the initial processes that exist within a hospital, it happens across the network as well. There's very interesting dynamics that happen on why it's set up a certain way. And if you want to talk about data cleanliness, people do some very strange stuff in EMRs with nested fields.

所以，尽管我们确实有电子健康记录和医疗系统，但它们各不相同，结构也很不规范，因此还需要对数据做一些转换。而且很多时候，提取数据很困难，所以到头来人们还是会选择发送传真或 PDF。因此，获取干净的数据并不是个容易解决的问题。So, even though we do have electronic health records and medical systems, they're all different, they're structured weirdly, so you do have to do some some transformations on it as well. And often times it's it's difficult to pull data out, so people just choose to send faxes or PDFs at the end of the day. So, solving clean data is not an easy problem.

嗯，AI 可以帮助我们，数字化转型也可以帮助我们，但在上游，有很多因素会影响这一切。嗯，这是一些重复记录的示例数据。嗯，实际数据也很相似，在某些网络中存在大量重复记录。如果你是这些重复记录的接收方，这确实会让人非常沮丧，对吧？所以，我们有不少客户会说：“这个组织能不能别再给我们发送这么多重复记录了？”对吧？除非你有完善的系统来识别这些重复记录，否则你就会不断收到同样的低质量数据。Uh AI can help us, digital transformation can help us, but upstream there's a lot of factors that go into it. So, uh this is an example of some illustrative data. Uh of duplicate records. Uh real data looks quite similar, as well, with a high number of duplicates across certain networks. And this is actually very frustrating if you are on the receiving end of these duplicate records, right? So, we have a number of customers who are like, "Can this organization stop sending us so many duplicates, right?" And unless you have good systems to detect this, you got you just keep getting the same bad quality data.

因此，拥有这些技术实现和检测系统非常重要，这样才能确保保护好自己的数据池，呃，数据的清洁度。那么，一旦有了这些数据，就有一些实际修复问题的机制，对吧？So, it's really important to have those technical implementations and detection systems to actually make sure that you can safeguard your own data pool uh data um cleanliness. Now, once you have this data, there there's mechanisms of actually fixing things, right?

出于监管或人际方面的原因，对吧？但这些杂乱的数据仍然作为基准存在。一旦你能检测出重复项之类的问题——这似乎已经是个解决了的问题——你就不再会鼓励人们多次提交理赔申请，每次都用不同的日期，或者提交同一份理赔申请；这样一来，其他人也不必再去清理他们的数据了。除了数据整洁度之外，另一点是，这些理赔申请的数据量确实很大。For regulatory or uh interpersonal reasons, right? But you still have this messy data that exists as a baseline. And once you can detect things like duplications, uh again, seems like a solved problem, uh you no longer incentivize things like sending claims multiple times with different dates or the same claim uh or just other folks have to clean up their data, as well. The other part that, beyond just the cleanliness of the data, is that these claims are really large.

呃，我刚开始工作时，在我们的数据库里看到有一个 8 GB 大小的 PDF 文件，这让我非常惊讶。Uh when I first started working and saw within our database that there was a PDF that was 8 GB in size, that was very surprising.

步骤 2：分析文档内部内容Step 2: analysis within documents

嗯，所以，这些理赔材料里真是什么都有。它们有图片，对吧？比如说车祸，对吧？里面会有受伤部位、碰撞现场等的照片，还有 X 光报告、超声检查，介于两者之间的各种材料都有。有时候，这些东西会让很多 PDF 阅读器直接崩溃。你想想那些老旧的电脑，要打开一个 8 GB 的 PDF 文件，那可相当困难。Um so, it these claims have anything and everything within them. They have uh images, right? Uh if you think about auto accidents, right? There are their images uh within the site of the injury, collisions, and such. There's x-ray reports, uh ultrasounds, everything in between. And sometimes this just crashes a lot of PDF viewers. Uh if you think about some of the old computers that you might have, trying to open up an 8 GB uh byte PDF will be will be quite challenging.

呃，甚至在技术系统方面也是如此，对吧？我们总是在谈上下文窗口。你要怎么把上下文塞进一个 8 GB 的 PDF 上下文窗口里，对吧？这并不像说“嘿，Claude，看看我的 10 页 PDF，然后给我个答复”，或者“嘿，ChatGPT”或“Gemini”，对吧？没那么容易。嗯，除此之外，实际上完整地处理一个案例也需要大量时间。Uh and even on the technology systems, right? We talk about context windows all the time. How are you going to fit a context uh into a context window in 8 GB PDF, right? It's it's not as easy as saying, "Hey Claude, take a look uh at my 10-page PDF and give me a response or hey chat GPT or Gemini, right? Not that easy. Um and on top of that, actually going through and working through a case requires a significant amount of time.

所以，要么是技术挑战，要么是导致非常复杂技术挑战的业务挑战。现在，如果我们逐步处理，并不断增加这里的复杂性，很多此类理赔也会存在不一致之处。比如，你可能会看到一些病情进展记录，其中较早的一份患者报告写着：“嘿，这是预约在某个日期拍摄的 X 光片。”但你收到理赔申请材料后，却发现里面没有这张 X 光片。So, either the technical challenges or the business challenges that result in very complex technical challenges. Now, if we work through and we're adding more and more complexity into here, a lot of these claims also have irregularities. So, you might have some progress notes, for example, of a earlier patient report that says, "Hey, here is this x-ray that was booked for this date." You receive the claimant package and that x-ray is not there.

对吧？如果没有以某种方式把这一点突出显示出来，就很容易忽略。嗯，再说一次，如果没有这一点，整个案件的走向都可能改变：是否提供了充分的护理、当事人是否真的接受了所提供的护理，以及你们如何将这些信息与账单方面交叉核对？Right? And it's very easy to miss that without actually having this highlighted in some sort of way. Um and without that, again, the whole trajectory of a case can change whether there was adequate care given, uh whether the person actually went to the care provided, uh and how do you actually cross-reference this with the billing side of things?

所以，这是一个非常非常常见的情况：由于上游的数据很混乱，你实际上不得不反复询问：“嘿，你能把这份记录发给我吗？”“哦，我们漏发了这份文件。”“哦，传真出故障了。”“哦，我们更换了后端系统。”对吧？因此，作为一个组织，你必须设置这些防线，以确保这些问题能被标记出来。而作为理赔员，他可能一次要处理 200 份理赔申请。So, this is a very very common scenario where upstream, again, because the data is messy, you actually have to ask repeated times, "Hey, can you send me this record?" "Oh, we missed faxing this." "Oh, the fax glitched." "Oh, we changed our back-end system." Right? So, with as an organization, you have to have these lines of defense in order to make sure that these get flagged. And as a claims adjuster, a claims adjuster might be processing 200 claims at a time.

所以，他们不可能把所有事情都记在脑子里，比如说：“哦，对了，我得跟进一下这张 X 光片，对吧？”他们可以做笔记、继续推进，但还是会有事情被遗漏。所以，这又是技术能够极大帮助解决这类问题的地方。接着，你会遇到相互矛盾的诊断结果。So, they're not going to be able to keep everything within their minds saying like, "Oh yeah, I need to follow up on this x-ray or not." Right? They can take notes, they can move forward, but things are going to get lost. So, this is again where technology can be a great lift uh to these kinds of problems. Then you get conflicting diagnoses.

这种情况也相当常见。那么，在座各位中，有谁曾因某种疾病寻求过第二诊疗意见？没几个人，对吧？嗯，有时候我们会去征求第二意见，听听另一种看法。有时候得到的答案完全不同，对吧？而索赔越复杂，出现某种分歧的可能性就越大。所以，在这种情况下，通常会发生的是：审核医疗索赔时，如果发现诊断结果相互冲突，或者想再次核实应该预留多少金额用于赔付，就会把案件转交给 IME，也就是独立医学评估。This happens fairly often as well. So, who here has sought a second opinion for a medical condition? Few folks, right? Um sometimes we seek them out um to get another perspective. Sometimes you get completely different answers, right? And the more complex the claim is, the more likely there is to be some some kind of conflict. So, in this case, what often happens is when you adjudicate a medical claim and you see some conflicting diagnosis or you want to double-check how much you should set aside to actually pay out for the claim, you send it to an IME or an independent medical evaluation.

接着还有另一位医生负责审核档案。然后你还得管理这些医生、医疗服务提供者网络，以及 IME 本身。所以现在，我们又在不断叠加了。Then you have another doctor who performs their review of the file. And then you have to manage those doctors and the provider networks and the IME themselves. So, now we're stacking again.

生态系统中的参与者越多，复杂性就越高。More people into the ecosystem, more complexity.

对医疗记录中的模型进行基准测试Benchmarking models on medical records

如今，AI 在这里确实能派上不少用场，而且我想，在 2026 年做演讲却不谈评估，多少有些不合时宜。不过，我们确实尝试用 AI 从某些方面解决这些问题，比如诊断结果相互矛盾、记录缺失，以及其他一些问题。因此，我们构建了一个名为 MLCR 的基准测试，借助人工分析来衡量 AI 实际回答这些问题的能力。现在，切换到我们刚才提到的博客文章，我往下滚动，看看这个基准测试是什么样的。Now, AI can be quite useful here and it would it would be, I guess, poor taste to give a talk in 2026 without talking about evals. But we did use AI to try to solve these issues in some ways, the conflicting diagnosis, the missing records and a number of other ones. So, we built this benchmark called MLCR with artificial analysis to measure the effectiveness of AI in actually answering these questions. So, flip over to our blog post that we talked about and I'll scroll down to what this benchmark looks like.

因此，我们发布了与医疗理赔相关的六个不同层级的问题。其中一些非常简单，比如这次受伤是什么时候发生的？伤情是什么？对吧？之后问题会变得稍微复杂一些：做了多少次手术？这位患者的能力如何？So, we released six different tiers of questions related to medical claims. Some of these are very simple, like when did this injury happen and what was the injury, right? Then they get a little bit more complex. How many surgeries get performed? How much capacity does the this patient have?

最终，我们会遇到越来越复杂、也越来越难以解决的问题。而且，作为理赔员、医生或管理人员，你提出问题时往往并不了解事情的完整背景。这时就会出现幻觉：人们以为病历中记载了某些内容，实际上却没有。好，我们再来看看结果。相关内容可在 Artificial Analysis 网站上查看。And eventually we get into more and more complex questions that are even harder to resolve. And oftentimes, the way that you ask a question as a claim adjuster, a doctor, an administrator, you don't have full context of what's happening. So, this is where we see hallucinations show up where people assume something is in a medical record, but it's not. So, let's go a little bit more to the results. So, this is available on the artificial analysis website.

如果你想看看，呃，我们发布了一个开源数据集，涵盖了我们问题的前三个层级，也就是比较简单的那些题。另外，我们还有一套私有数据集，由 Artificial Analysis 托管，目的是在一定程度上防止针对基准测试进行优化。所以，看看模型的表现非常有意思。呃，实际上，我们会从三个不同的维度来衡量这一点。If you want to take a look, uh, so we released an open source, uh, data set for the first three tiers of questions that we have, the easy ones. And we have a private set that we host with artificial analysis for, um, to stop benchmarking to a certain extent. So, it's very interesting to see how how models perform. Uh, we actually measure this on three different, uh, axes.

那么，我切回那张幻灯片。我们会衡量简洁性，也就是回答有多简洁。我们会将回答与经过人类专家验证的答案进行比较。接下来是准确性：答案正确吗？最后一项是完整性：问题中所有相关信息是否都得到了回答？那么，我们回到这里。前沿技术仍在以非常非常快的速度发展，这正是评估至关重要的原因。So, let me flip back to that slide. We measure this on conciseness. So, how concise is a response? Uh, and we measure this against a human, uh, expert validated answer. We then go through accuracy. Is the answer correct? Uh, and the final one is completeness. Are all the relevant, uh, informations, uh, piece of information answered, uh, in the question? So, we go back here. Um, the frontier has continued to evolve very, very quickly, and this is where having evals is really important.

呃，我们发布这项评测时，呃，在我们设置的推理难度等级中，观察到的最高准确率是 20%。几个月后，在最高难度的推理等级上，我们已经达到了 75%。对吧？所以，模型进步得非常快，嗯，但这是件好事，因为这意味着我们可以开始将许多更复杂的问题产品化。Uh, when we released the eval, uh, the highest accuracy that we saw on the levels of reasoning that we had was 20%. A few months later, we're in the 75% with the very high reasoning levels. Right? So, models move very quickly, um, but this is great because it means that we can start productizing a number of the more complex questions.

而且，我们更确定有人需要在系统中进行审核。嗯，在我们的系统中，始终有人参与其中。这一点非常重要。你也开始看到帕累托前沿上的成本曲线，也就是提出这个问题并获得合适答案的实际成本。过去几个月，这方面也一直在改善。你还会看到其中的一些权衡。And we have more certainty of the confidence that somebody has to review this, uh, in a system. So, within our system, there's always a human in the loop. Uh, so that's really important. You also start getting the cost curves here on the Pareto frontiers, on the cost to actually ask this question and get a suitable answer. And this has been improving through the past few months as well. And you get some of these tradeoffs as well.

因此，我们在对文档分析进行这些基准测试时发现一件有趣的事：GPT 模型的准确率非常高，但遗漏了一些关键信息。Anthropic 的模型表现也相当不错，而且提供的信息要完整得多。不过，它们的准确率也略低一些。So, one of the interesting things when we were running these benchmarks on, uh, document analysis was that the GPT models were excellent at accuracy, but they omitted some key pieces of information. Uh, Anthropic's models were quite good, um, and they're much more complete on the amount of information. Um, they do have slightly lower accuracy here as well.

观众：我可以问个问题吗？Audience member: Can I ask a question?

演讲者：当然。Presenter: Sure.

观众：准确性，你们如何衡量它与人为因素的关系？Audience member: Accuracy, how do you measure it compared to the human factor?

演讲者：是的，与人类答案相比，然后我们会进行三重评审验证。嗯，是的。这里展示的是准确性，你可以看到曲线的走势；还有完整性，这一问题在大多数模型提供商那里仍未解决。至于之前的问题，是的，这基于经过人工核验的答案，以及一组经过校准的评审模型。在我们的博客文章中，我们介绍了这里采用的方法，之后也发布了一篇后续文章，说明我们是如何进行校准的。Presenter: Yes, compared to a human answer uh and then we do the triple judge uh verification. So, yeah. So, we have accuracy here. You can see how uh the curve looks like and we have completeness uh which is uh is still not solved across most model providers. And yes, to the earlier question, uh this is based on a human verified answer and on a calibrated set of judges. Uh within our blog post, we describe the methods we have here uh and we released a follow-up as well on how we did this calibration.

所以，尽管我们在这里使用 AI 来整理数据并提取重要见解，但我们要确保这个过程的每一步都经过良好校准，因为如果上游的数据有误，而你又把它用于下游，对吧？那你其实什么也没得到。所以，再强调一次，在推进理赔流程时，校准好评判模型、构建优质评估非常重要。So, even though we're using AI here to structure our data and extract important insights, we want to make sure that each step uh of that process is well calibrated because if upstream uh your data is incorrect and you start using it downstream, right? You haven't gained anything. So, again, really important as you're going through the claims uh uh process to calibrate your judges, to build good evals.

嗯，我觉得这是 Ed Most Talks 一直反复提到的主题。你要确保有能够真正核实这些信息的专家，因此需要跨团队协作来实现这一点。好，这些都是针对单笔理赔的。呃，我们知道，只要有一定程度的把握，就能提取某些类型的洞见。我们可以标记相互冲突的诊断，可以去重，也可以查找缺失记录。接下来聊点更有意思的，也就是跨理赔。Uh I think that's a recurring theme at Ed Most Talks. And you want to make sure you have those experts who can actually verify this information, so you work across teams to get there. Okay. So, that's within a claim. Uh we know that with some amount of certainty, we can extract some types of insights. We can flag conflicting diagnosis. We can deduplicate. We can search for missing records. Let's go to something more interesting, which is cross claims.

步骤 3：跨声明的洞见Step 3: insights across claims

对吧？所以，如果你是一家中型保险公司，业务涵盖多个险种，你的保费收入可能达到 10 亿美元。你每年可能要赔付大约 2 万笔理赔，甚至更多，具体取决于你经营的险种。现在，你想更深入地了解究竟是什么因素影响了你的赔付率。人们提出的都是哪类理赔？这些理赔发生在哪些地区？Right? So, if you are a mid-size uh insurance company, you might be collecting $1 billion in premiums within a uh portfolio of lines of businesses. And you might be paying out, let's say, 20,000 uh claims uh at a time per year uh or even more, depending on what line of business you deal with. Now, you want more insight of where you actually like what's affecting your loss ratio. What kind of claims are people making? In which geographies?

是什么类型的伤病？涉及哪些类型的医生？我认为这也是很有意思的地方。而且，一旦从索赔数据中提取出这些信息，就可以开始进行有趣的可视化、回归分析等工作。对吧？我们被问到并着手回答的两个问题之一是：人们在一个网络内流动的情况如何？What type of injuries? and what type of doctors are being involved. This is where I think it gets quite interesting as well as actually being able to, once you have that data extracted from the claim, you can start doing interesting visualizations, regressions, and work as well. Right? So, two of the questions that we've been asked and we've worked on answering is sort of how much are people traveling within a network, right?

所以，如果患者比如说在旧金山或纽约，对吧？他们是在当地接受治疗吗？提供治疗的人选对了吗？我们还可以查看这些人员之间实际合作的关系网络。这里展示的是一部分经过隐去处理的加州数据，你可以追踪哪些律师在合作，哪些医生在合作，哪些脊椎按摩师、物理治疗师在合作，以及实际的合作关系是什么样的，还有理赔本身的情况。So, if a patient is, let's say, in San Francisco or New York, right? Are they getting treatment in their local area? And is that the right person to be giving the treatment? We also have these networks of who's actually working together. So, this is some redacted data, a subset from California, and you can actually track to see which lawyers are working together, which doctors are working together, which chiropractors, physiotherapists, and what is the actual working relationship look like and the story within the claim itself.

所以，这是从你的理赔数据中提取信息时能够获得的一类更高层次的洞察。接下来就会涉及一些有意思的医疗结果，对吧？比如，哪些病症会同时出现？举个例子，如果有人背部受伤，对吧？So, this is that sort of higher-tier insight that you can get within extracting data from your claim. You then get into some interesting medical outcomes, right? So, which conditions co-occur together? So, if somebody has a back injury, as an example, right?

你可能会有一系列不同的诊断，它们可以映射到 ICD 编码并同时存在；你可能会想，为什么某位医生的记录和另一位医生的记录会出现这种差异？在这种情况下，还是那句话，只有解决了单个理赔中的数据提取问题，并将其汇总到更广泛的保险组合中，你才能实现这些能力。现在，我们来谈谈费用部分，好吗？比如，这里取一个很小的子样本，包含两项物理治疗，不同医生收取的费用各不相同，对吧？You might have a series of different diagnoses that could be mapped to ICD codes that coexist together, and you might see, okay, why does this exist from a certain doctor's write-up versus another's? And in this case, again, you only get these capabilities if you can solve the data extraction problem within one claim, and then roll it up into your broader insurance portfolio. Now, we get to the money portion, right? So, pulling a small subsegment here of two treatments of physiotherapy, different doctors charge different amounts, right?

我们大概都见过这种情况。有时候这也没什么问题。如果你是更好的医生，就应该能收更高的费用，对吧？如果你能让患者尽可能快地康复，只用五次治疗而不是十次，就应该能收更高的费用，对吧？We've probably seen that. And sometimes that's fine. If you're a better doctor, you should be able to charge more, right? If you can get somebody to recover as quickly as possible in five treatments instead of 10, you should be able to charge more, right?

所以，呃，理解成本并专注于尽可能降低成本真的非常重要，因为医疗体系的不同环节一旦出现成本通胀，就会不断累积，对吧？保费会上涨，呃，我们被收取的费用也会上涨，呃，而我们花同样的钱却能得到更少的服务。所以，在这项分析中，呃，你可以看看各项理赔，是否存在一个一致的因素？你应该提供多少保障？So, uh but it's really important to understand costs and focus on minimizing them because as soon as there's cost inflation through different parts of the medical system, it keeps compounding, right? Premiums increase, uh the fees that we are charged increases, uh and we get less for our money. So, within this analysis, uh you can see across claims, is there a consistent factor? What what is the amount of coverage that you should be providing?

那么，理赔时这些费用应该如何核保？你也可以估算这些费用会如何随时间变化，对吧？比如说，按年计算，某位医生提供的这类治疗费用上涨了 10%，对吧？这种情况会在整个司法管辖区内发生吗？还是会在整个行业普遍发生？How How you should you be underwriting these costs for claims as well. And you can estimate how things are going to change over time, right? So, year over year, as an example here, uh there was a 10% increase for this type of treatment for a given doctor, right? So, is that going to happen across uh the jurisdiction? Is that going to happen across the industry in general?

那么，你们就可以开始做一些反事实分析，或者在分析中进行一些调整，比如说，如果我们改变网络，如果我们开始引导人们朝某个方向就医，如果我们改变计费方式，对吧？可能节省多少？根据现有的医疗护理标准，人们应该接受哪些治疗程序？嗯，最近有篇文章谈到过度计费，有人看到那篇文章了吗？文章说，额外的账单金额达到了 10 亿美元。So, then you can start doing some counterfactual analysis or changes in a in analysis to say, if we change our network, if we start directing folks into a certain direction, if we change how we how we bill, right? What are the potential savings that exist? And what kind of procedures should people be having based on the standard of care that exists? Um there's recently an article about how as anybody see it seen at the overbilling article about $1 billion of additional billings that have come in?

你看到那个了吗？嗯，这些都是例子，再次说明，如果你无法有效处理数据，人们就会想办法绕过流程、提交额外记录，或者做一些更可疑的事情。总而言之，有了良好的数据之后，你就可以开始做出这些决策。Have you seen that? Um so, those are examples where again, if you cannot process data effectively, people are going to find ways to go through and submit additional records uh or do some some shadier behavior as well. So, all this to say is that once you have good data, you can start making these decisions.

然后，一旦你掌握了这个决策——呃，这些理赔信息，就可以开始把它写入一个更广泛的系统，对吧？这样，你就可以开始将这些信息用于其他业务线，比如说，如果你知道汽车事故伤害或工伤赔偿方面有哪些趋势，我们在一般责任险中是否也看到了同样的情况？如果你知道承保模型遗漏了某些趋势，或者它的可解释性不够强，我们能否根据观察到的情况，让模型变得更具可解释性？And then once you have this decision, uh this claims information, you can start writing it into a broader system, right? So, you can start using this in the other lines of business of saying, if you know these are the trends in auto injury or workers' compensation, are we seeing the same thing in general liability? If you know that your underwriting model is missing certain trends or you don't have as much explain ability there, can we make the model more explainable with what we're seeing?

我们如何在数据湖中提供这种我们希望存在的神奇数据表？再强调一次，数据越多，所能构建的图景就越完整，最终也就越能为消费者准确承保并合理定价。而这些信息正是推动高效 AI 应用所必需的，无论这些应用用于承保、欺诈检测还是理赔。How do we provide this sort of magical data table that we wish that existed within our data lake? So again, the more data you have, the more complete of a picture you can make and the better you can underwrite and price the product at the end of the day for the consumer. And this is information you need to actually power the effective AI applications, whether they're on the underwriting side, on the fraud side, on the claim side as well.

所以，如果你想象一下一个理想化的世界，对吧？我们希望从更广泛的层面减少某些东西。比如说，如果我们想为所有人降低成本，同时考虑到患者体验；如果我们想缩短理赔未结案的时间；如果我们想让人们更快重返工作岗位，并改善他们接受的治疗，那么我们就必须确保拥有正确的数据，并进行恰当的数字化转型。So if you think about like a blue sky world, right? We want to reduce something more broadly. So if we want to reduce costs for everybody, right? Taking into account that patient experience. If we want to reduce the time that a claim is open. If we want to improve how quickly people get back to work and the treatment they get, we need to make sure we have the right data and the right digital transformation to do that.

这也是我们在 WiseDocs 工作中所做的一个说明性示例。我们也看到了显著的成本节省。因此，推进这一数字化转型流程并拥有正确的数据，确实非常重要。我就到这里结束本次研讨会。我想留出充足的时间来回答问题，因为这是一个相当全面的话题。感谢大家聆听，欢迎提问。So this is an illustrative example in the work that we've done within WiseDocs as well. We've seen significant savings as well. So going through this digital transformation process and having the right data is really really important. So I'll wrap up the workshop here. I did want to leave quite a bit of time for questions cuz it is pretty comprehensive topic. So thank you for listening and happy to take some questions.

是的，我会、我会分享这篇博客文章。是的，这边有个问题。嗯，问题是，我们到底该如何组织数据，来构建所有这些信息。它会经过多个不同的层级。我也会分享一些我围绕这个主题做过的演讲。Yes, I'll I'll share the blog post. Yes, question over here. Yeah, so the question was about how do we actually structure the data to build um all this information. So it goes through a number of different levels. I'll share some talks that I've given on the topic as well.

处理栈The processing stack

基础做法是使用 OCR 和图像从底层记录中提取信息。这算是一种标准的预处理。如果有人处理过 PDF，就会知道预处理 PDF 也不是件容易的事。呃，那些 8 GB 的文件里会有一些非常罕见的奇怪案例。所以，拿到数据，也就是文本和图像之后，我们会在此基础上使用一系列模型，比如 NLP 模型和信息提取模型。有时我们会用 VLM 来构建结构化数据；像更标准的实体提取模型也很有效。The baseline is we use OCR and images to extract information from the underlying records. Sort of standard pre-processing. Pre-processing PDFs is not an easy task as well if anybody has worked with them. Uh those 8 gigabyte ones have some some weird one in a million examples in them. So, once you have the data, the text and images, we have a series of models on top, um NLP models, extraction models. We use VLMs uh to structure it sometimes, like a more standard entity extraction model works well.

我们会在此基础上构建各种分类器。嗯，然后再在上面加一层摘要层，用来总结文档。接着由人工核验，确保信息准确。最后，我们会在此基础上运行其他分析层。因此，我们采用了一种抽象模式，随着层级上升，信息密度也逐渐集中；不过，在不同时间点，我们会通过各种模型和人工干预来进行核验。We have various classifiers that we build on top. Um we then have a summary layer that sits on top of that to summarize the documents. Uh that is then human verified to make sure that the information is correct. And finally, we run those additional layers of analysis on top of that. So, we have this abstraction pattern that sort of concentrates the density of information as you go up, um but it's verified at various points in time with various models and human intervention as well.

所以，这意味着当你在最有价值的项目中做到顶尖水平时，AI 才能发挥作用。而这类问题大致属于 MLCR 类型，能够有明确的正确答案。在 MLCR 中，实际上是使用干净的数据来提出这些问题。现实世界的数据并不干净，因此在此之前必须经过多个步骤来清理数据，才能达到这一点。是的，我们的客户通常——规模较大的客户内部有多条业务线，而且我想说，不同团队之间的数据共享规则各不相同。So, that means once you get to the top in the highest value items, that's when AI can be effective. And those are sort of those MLCR uh type questions that can have the right answer. So, within MLCR, this actually uses clean data to ask these questions. Real-world data is not clean, so you have to have various steps ahead of that to clean up the data to get there. Yeah, so typically our customers So, bigger customers have multiple lines of business internally and they have different uh rules I would say across sharing across different teams.

嗯，其他客户只在一个业务领域开展业务。所以，他们可能只做汽车保险，也可能只做残疾保险，或者只做工伤赔偿。不过，我们会综合各方面的趋势，尽量识别并标记出这些情况。Uh other customers just work in one line of business. So, they might just work in auto, they might just work in disability, they might just work in workers' comp. Um but we take the trends uh across the board to try to figure out and flag those.

所以，问题是，在某个办公室内部，如何呈现其中一些洞察，或者说地理信息？然后是上面一层，可能是更广泛的公司层面；不过，保险公司有时也会包含一系列独立运营的子公司。然后，如何将这些数据实际汇入一个集中式数据湖？所以，同样，每家公司都有些不同，但我们会按照合作伙伴希望的速度推进。So, there's the question that within sort of an office, right? How do you surface some of those insights uh or sort of geography? Then you have the layer on top, which might be um the broader company, but comp- insurance companies also have a collection of companies within them sometimes uh as separate businesses. Uh and then how do you actually feed that into a centralized data lake uh as well. So, again, every company is a bit different, but we go as fast as our partners want us to.

我们深耕理赔领域，因此专注于理赔业务。理赔是我们的核心，尤其是复杂理赔，我们整个 AI 流程都围绕这类业务展开。但随着我们不断发展，并逐步解决理赔裁定、理赔处理和信息提取等相对简单的挑战，我认为，我们开始发现一些更具洞察力的方面，能够帮助机构在业务组合层面做出决策。We live in claims. So we focus on claims. So claims are our bread and butter, specifically complex claims. That's where our whole AI pipeline is focused on. But as we sort of evolve and go and solve the sort of easier challenges, I would say, with adjudicating a claim, processing a claim, extracting that information, you start seeing these more sort of insightful aspects that help organizations make portfolio level decisions.

是的，是的，这是个好问题。问题是，面临这么多挑战时，我们该如何让某人加入并熟悉情况？通常，我们合作时最看重的是改变的意愿，对吧？我们大概都和想要改变的团队或组织合作过，也和不想改变的合作过。嗯，体验非常非常不同。假设这是一个想要改变的组织，我们通常会用几份示例文件，分别以不同格式来开展工作。Yeah. Yeah, so that's a good question. So the question is how do we onboard somebody when there there's a lot of these challenges? So typically we work with the most important thing is a willingness to change, right? So we've all probably worked with teams or organizations who want to change versus those who don't. Um very very different experience. So assuming it's an organization that wants to change, there'll be a few sample files typically that we'll work through in various formats.

通常，人们会把他们正在处理的内容、手头已有的内容，或者基本上想要进行回测的内容发给我们。然后我们会进行审核，看看你提取了什么、我们的系统提取了什么，两者是否一致？它是否符合你想要查看的信息？你们有不同类型的标准吗？Typically people will send us either things they're immediately working on or have or are basically they're trying to back test. Uh and then we go through a review process to see what did you extract, what did our system extract, how does it align? Does it match the information you want to see? Do you have different types of standards?

所以我们回过头来，比如说，呃，看看这里的不同类型的文档，我们就回到受理表单吧。这份表单上可能有多个日期。你想要的是受伤日期，还是提交时的报告日期，对吧？这些都涉及业务目标的协调。先用几个标准案例走一遍流程。如果进展顺利，我们就开始更大范围的试点，然后进入有趣的集成环节：除了手动上传文档，还要与上游服务提供商集成、对接 API，处理各种各样繁杂的问题。So if we go back, for example, uh to the different types of documents here, let's just go back to the intake form. There might be various dates on this form. Do you want the date of injury or do you want the report date when it was submitted, right? So these are things of alignment of business objectives. Go through process a few standard cases. If things work well, then we start sort of a broader pilot and then gets into the fun integration aspect of integrating with upstream providers apart from just like hand uploaded documents, API integrations, like dealing with all all the messiness.

因此，只要有明确的业务成果和转型目标，整个流程就相当直接。当然，在推进这一过程时，有时我们现成的能力还不足以满足需求，有时则可以满足；有时还需要与多个内部系统进行集成。So as long as there's a clear business outcome and transformation goal, then it's it's a fairly straightforward process. And again, as we go through that journey, and sometimes it's like we don't have the capabilities out of the box, sometimes we do, sometimes there's multiple integrations on internal systems as well.

嗯，是的。我们先从问题的第一部分说起，也就是如何应对监管要求，以及我们如何与合作伙伴合作。这取决于人们行动的速度，也取决于监管变化的速度。我们目前的系统设计是，始终需要有人参与其中做出决策。比如，对于诊断结果相互矛盾的情况，我们会把这些洞见呈现出来，告诉你：“嘿，第2页……第20页上，这两位医生的诊断结果不一致，你应该看一下，对吧？”所以，只要有人参与决策，我认为就能消除人们大约70%的监管方面的顾虑。So, yeah. So, let's start with the first part of the question, which is handling regulations and how do we work with partners? So, we it depends on how how quickly people are moving. Uh and also depends how quickly regulations are moving. So, the way we've set up the system right now is that it always requires a human in the loop to make a decision. Uh so, we surface these insights as an example for things like a conflicting diagnosis, we'll see, "Hey, on page two on page 20, there there's a difference of diagnosis between these two doctors, you should take a look, right?" So, as long as there's a human in the loop, I'd say it removes like 70% of the regulatory concerns that people have.

嗯，监管方面也还在跟进，所以有很多资料可以阅读，比如我提到的 NAIC 报告。该组织会发布许多不同的报告。Uh regulation is also still catching up, so uh there's a lot of things to read about like the NAIC reports that I mentioned, they that organization publishes a number of different reports.

对此已有相关指导方针。我们也会与内部治理团队和监管团队沟通。如果他们想了解一些信息，我们会向他们说明 AI 流程是如何运作的，对吧？使用的是哪类数据。但归根结底，只要不构建黑箱，确保整个过程可审计，并且有人参与其中，就能解决很多问题。There's the guidelines that exist for that. We talk to the internal governments and regulatory teams as well. Uh and if they want to see some pieces of information, like we will share with them like this is how the AI pipeline works, right? This is the type of data that's being used. Um but at the end of the day, as long as you don't build a black box and things are auditable uh and there's a human in the loop, that takes care of many things.

在监管方面，还有另一整套业务决策层面的考量，比如如何应对核弹级判决。呃，核弹级判决指的是金额非常大的赔偿裁决或索赔，金额达到 1,000 万美元；而热核级判决则达到 1 亿美元，这会大幅重新分配企业的准备金。所以，对于这类案件，人们通常会提出自己想如何处理的方案。The regulatory side, there's like a whole other like business decision aspect, like how to approach nuclear verdicts. Uh for uh nuclear verdicts are um very large uh awards or um claims that are $10 million or thermonuclear ones which are $100 million, which really reallocate uh businesses' reserves. So, for those ones, typically people come with like an approach on how they want to handle them.

在平台内，你可以说：“我要先处理最复杂的理赔申请。”或者“我要先处理最早提交的那笔申请。”嗯，这些都是人们可以自行决定的业务事项。呃，关于问题的第二部分，也就是人们或组织内部的目标。嗯，我的意思是，我们希望与那些抱有相同理念的合作伙伴携手合作：改善患者护理、减少理赔积压，并做出最佳决策。And within the platform, you can say, "I'm going to handle my most complex claims first." Or I'm going to handle the one that was submitted first. Um and those are all sort of business decisions people can make. Uh the second part of the question on like people's internal like organizations' internal goals. Um I mean, we want to work with partners that share the same sort of mindset about improving patient care, uh reducing backlogs uh for claims, providing the best decisions.

很多时候，我们的客户甚至不太会去想医疗本身的费用，他们在意的是围绕医疗的一切，对吧？A lot of times our customers are not even thinking too much about like the cost of health care itself is everything around it, right?

监管、成本与未来展望Regulation, cost and what's next

所以，呃，如果你是一家大型保险公司，实际上会在法律费用上花费数亿美元，对吧？这件事甚至和患者体验无关，但在这方面有很大的优化空间，也可以加快处理速度，对吧？嗯，还有监管方面的顾虑。在加利福尼亚州等某些司法辖区，诉讼非常频繁。嗯，你会想弄清楚该如何应对这些问题。So, uh if you're a large insurance company, you actually spend hundreds of millions of dollars on legal fees. Right? So, that is something that is not even tied to the patient experience that there's a large amount of optimization you can do there and to speed things up. Right? Um there's regulatory concerns that in certain jurisdictions like California, things are very litigated. Um you want to figure out how how to deal with those.

嗯，还有一大堆中间方也参与其中，而且他们也效率低下。所以，视客户而定，每个人的目标都不一样。不过在我们的对话中，我们始终觉得，我们有一个非常明确、可以用技术解决的问题，那就先把它解决。然后，相关洞见就会陆续浮现出来，比如：“哦，我以前都不知道我的业务还有这种情况。”Um there's a whole bunch of middle parties that are involved as well uh that are also inefficient. So, depending on the customer, everybody has a different goal. Um but in our conversations, it's it's always been like we have this very obvious problem to solve with technology. Let's solve that first. Um and again, the insights start surfacing afterwards of being like, "Oh, I didn't know that about my business.

“那我们之后再解决这个问题。”所以，是的，问题是关于 Jv 中的决策 API。嗯，作为……我不知道有没有人看过会议发放的《AI News》。我正在为它写一篇文章。嗯，我们长期以来一直在使用分类模型。我通常对此有很强烈的看法。这篇文章今天不会刊出，会在之后的一期刊出。Let's solve that afterwards." So. Yeah, so the question is about decision APIs in Jv. Um as part of the I don't know if anybody's seen the paper um copies of the AI news that are that the conference is handing out. There's actually an article that I'm writing for that. Um so, we've been using classification models uh for a long time. I've generally had uh strong opinion on that. Hasn't come out today. It's coming out in a later edition.

嗯，我觉得这非常有用，不过大型语言模型并不是所有问题的正确解决方案，对吧？我们之前讨论过这个，对吧？这部分 AI 也可以进一步细分，对吧？比如分类、抽取式任务，以及相对更复杂的、基于决策的任务。嗯，对我们来说，成本非常重要，所以我们也训练了自己的模型。准确性也非常重要，还有幻觉问题。Um I think it's very useful, uh but LLMs are not the right solution for everything, right? So, when we talked about this, right? This segment of AI can be split up as well, right? Classification, extractive uh versus sort of the more complex decision-based ones. Uh so, for us cost is really important. So, that's why we trained our own models as well. Accuracy is really important and hallucinations as well.

所以，我觉得 Jv 对它问世这件事并没有那么惊讶。更让人惊讶的是它花了这么久才出现，因为我觉得，LLM 对每个问题都进行 token 上限推理的这种现象持续了很长时间，直到今年我们才开始更多地考虑成本。所以，呃，Javis 已经上线之后，你就可以试用了。嗯，不过我们也一直在按照同样的思路构建自己的模型。就是这样。So, I think Jv was not as surprised that it came out. It was more so a surprise of how long it took because I think the LLM like token max reasoning for every problem phenomenon lasted for a very long time until we started thinking about costs a lot more this year. So, uh once Javis sort of hit already, we could you can test it out. Um but we've sort of built our own models with the same paradigm in mind. So.

演讲者：是的，我来回答第二部分。嗯，作为一个组织，我们推动整个行业变革的程度，还没有达到我们期望的水平，对吧？所以，我们不能说已经把每位患者的理赔时间平均缩短了 5 天。我们很希望能这么说，但目前还做不到。对于一些特定的组织，我们在网站上有一些案例研究。我也很快放上了一些。其中有些案例在线业务领域取得了显著成果，降低了成本，也提升了速度。Presenter: Yeah, so I'll answer part two. Um so as an organization we have not moved the broader industry um to the extent we want, right? So, to say that we can say we've reduced the claim uh time by 5 days on average for somebody across every patient. We'd love to say that. We're not in that position yet. For specific organizations like we had some case studies on the website. I quickly put put some up as well. There's some significant ones within lines of business uh for cost reduction and speed.

嗯，你可以看看那些内容；如果你还想了解更具体的背景，我们之后也可以聊聊。Um so, you can check those out and we can chat afterwards as well if you're looking for more specific sort of context on that, too.

演讲者：是的，所以首先是伤害或损失通知，嗯，具体还要取决于子领域。是的，嗯，还要看司法辖区，在处理这些事项时也有时限要求。比如，在德克萨斯州，你只有几个小时来处理索赔并给出答复。在其他州，你可能需要在 7 天或 21 天内作出答复。现在情况会变得更复杂，因为如果你是一名在工厂工作了 30 年的工人，假设你已经 60 岁了，那你就会有一大份医疗记录档案，对吧？Presenter: Yeah, so the first notice uh of injury or loss um or and depend depending on the on the sub domain. Yes, um so depending on the jurisdiction there's also turnaround time requirements when handling these. So, for example in Texas you only have a few hours of processing a claim to get an answer. In other states it might be 7 days, 21 days that you need to get a response. Now, it gets more complicated because if you're a worker who's been working in a factory, let's say for 30 years, you're 60 years old, you're going to have large medical record file, right?

所以，这里以及初始报告中包含很多信息，但要进行全面的理赔分析，你得回头试着把这些信息都找出来。也许你结婚后改了名字，也许你搬了家，换了城市或服务提供商，对吧？So, there's a lot of information here and in the initial report, but to do a full claim analysis you have to go back and you have to try to pull those. Maybe your name changed because you got married. Maybe you moved uh cities, providers, right?

所以，你还得把那段完整的历史也大致重建出来。我觉得遗憾的是，在很多诉讼案件中，有些信息也会被故意遗漏，因为很多组织尚未完成数字化转型，无法足够快速地处理案件。所以，如果只需要这些信息，处理起来会快得多；但遗憾的是，或者更准确地说，实际情况是，背后的来龙去脉要复杂得多。So, you have to reconstruct sort of that full uh history as well. Uh and I think unfortunately in a lot of litigated cases there's information that's left out uh as well on purpose to due to the fact that a lot of organizations have not digitally transformed and can't process cases fast enough. So, uh if this was all that's needed, this would be uh much faster, but unfortunately or more realistically, there's a much broader backstory as well.