为什么AI终于改变了医疗保健中欺诈、浪费和滥用的数学
Why AI Finally Changes the Math on Fraud, Waste, & Abuse in Healthcare
美国医疗保健系统损失超过每年1000亿美元因欺诈、浪费和滥用，且据一些估计这个数字还要高出数倍。这一点我们已经知道了几十年。新的变化是，识别这些问题的技术终于跟上了。问题不再在于这些工具是否有效，而在于处理并支付有问题索赔的机构是否会最终使用它们。
系统实际损失了多少
三个不同的问题被混为一谈，因此有必要将它们区分开来。
欺诈
是有意欺骗：为从未提供的服务开具账单、幽灵患者、回扣团伙。
National Health Care Anti-Fraud Association估计仅欺诈一项就占医疗支出的3%（低端）至10%（高端）。相对于4.9万亿美元的全国医疗支出，即使按低端计算，也约为每年1500亿美元。
浪费
是过度利用：不必要的检查、重复的诊疗项目、增加成本但未改善护理的服务。
滥用
介于两者之间。它是不当计费，例如编码升级和拆分计费，这会抬高支付金额，但可能达不到刑事欺诈的标准。
2024年，Justice Department从医疗欺诈中追回约17亿美元。与1500亿美元的损失相比，该系统追回的钱大约只有每美元一美分。
为什么旧的策略不起作用
这么低的追回率是结构性的。健康计划先支付，后调查。只有在索赔提交、裁决并支付之后——这一过程可能需要数天甚至数周——才会有人检查它是否合法。目前主导这种事后审查的有两种工具，而且两者都有同样的缺陷：
规则引擎
会标记出符合已知模式的索赔，比如皮肤科医生在一天内为200项诊疗服务计费。它们只能发现人们已经定义过的骗局，所以总是慢人一步。
回顾性审计
对已经结清的索赔进行抽样。当审计人员整理案件时，提供者往往已经多开出数百万美元的账单，或者已经搬迁、停业。
这两者还会产生大量误报。调查人员最终把大部分时间花在澄清合法索赔上，而不是追查真正的欺诈。
AI有何不同
新颖的AI技术完全颠覆了这种方法。AI不再匹配已知模式，而是学习正常情况是什么样子，并标记出偏离之处。
三项关键的AI能力最为重要：
异常检测
模型学习特定专科、地区和患者构成的正常计费情况，然后突出离群值：编码强度远高于同行的提供者，或利用率曲线朝错误方向弯曲。一项2025年综述发表在Artificial Intelligence in Medicine上，研究发现难点已不再是发现异常，而是在欺诈只占数据极小比例且调查人员需要信任结果的前提下进行发现。
网络分析
代价高昂的骗局是团伙而非个人：医生将患者引导到他秘密拥有的实验室，药房为同谋处方者开出的处方配药。图模型可以绘制提供者、患者和设施之间的关系，并揭示任何逐项审查都无法发现的群体。
文档审查
语言模型可以读取临床记录，并将其与计费代码进行比较。这就是浪费和滥用所在：文件记录不支持所收取的服务级别。
真正改变经济算盘的能力是预支付评分：在索赔支付之前让这些模型处理它们，从而先扣下或审查可疑的索赔。这将整个问题从追回转变为预防。不再事后追捕欺诈者，而是根本不会付钱给他。
未来的机遇在哪里
这里的获胜初创公司将不是销售资金流失后略微改进审计的公司。它们将是把检测嵌入支付流程的公司，而更困难的是，连接多个支付方数据的公司。跨支付方数据将让这些公司构建持续改进的智能体系统，每天都能逐步更好地发现欺诈。
这里需要的最后一步是大规模采用：让大型支付方等规避风险的机构在错误指控会带来真实后果的环境中根据概率性标记采取行动，并让它们共享它们有充分理由囤积的数据。这个问题比发明技术容易得多，而我们已经完成了困难的部分。唯一悬而未决的问题是，医疗行业还会把千亿美元的漏损当作做生意的成本多久。
关于Kasra Khadem
Kasra Khadem是Pathlight Ventures的合伙人，这是一家位于纽约市的早期风险投资公司，他负责领导医疗技术投资。在加入Pathlight之前，Kasra曾在Human Capital担任投资者，与Commure、Transcarent和Ambience Healthcare等领先的医疗初创公司合作。
The U.S. health care system loses more than $100 billion a year to fraud, waste, and abuse, and by some estimates several times that. We have known this for decades. What is new is that the technology to catch it has finally caught up. The question is no longer whether the tools work; it is whether the institutions processing and paying the faulty claims will finally use them.
What the system is actually losing
Three different problems get lumped together, so it is worth separating them.
Fraud
is intentional deception: billing for services never rendered, phantom patients, kickback rings. The National Health Care Anti-Fraud Association estimates fraud alone at 3 percent of health spending on the low end, and as high as 10 percent. Against $4.9 trillion in national health spending, even the low number is roughly $150 billion a year.
Waste
is overutilization: unnecessary tests, redundant procedures, services that raise cost without improving care.
Abuse
sits in between. It is improper billing, like upcoding and unbundling, that inflates payments but may not meet the bar for criminal fraud.
In 2024, the Justice Department
recovered about $1.7 billion from health care fraud. Set that against $150 billion in losses and the system gets back roughly a penny on the dollar.
Why The old playbook fails
This low recovery rate is structural. Health plans pay first and investigate later. Only after a claim is submitted, adjudicated, and paid – a process spanning days if not weeks – does anyone check whether it was legitimate. Two tools currently dominate that after-the-fact review, and both share the same flaw:
Rules engines
flag claims that match a known pattern, like a dermatologist billing 200 procedures in a day. They only catch schemes someone already thought to define, so they are always one step behind.
Retrospective audits
sample claims that already cleared. By the time an auditor builds a case, the provider has often billed millions more, moved, or shut down.
Both also produce huge volumes of false positives. Investigators end up spending most of their time clearing legitimate claims instead of chasing real fraud.
What AI does differently
Novel AI techniques flip this approach altogether. Instead of matching known patterns, AI learns what normal looks like and flags what deviates.
Three critical AI capabilities matter most:
Anomaly detection.
Models learn normal billing for a given specialty, region, and patient mix, then surface the outliers: the provider whose coding intensity is far above peers, the utilization curve bending the wrong way. A 2025 review in Artificial Intelligence in Medicine found the hard part is no longer spotting anomalies. It is doing so when fraud is a tiny fraction of the data and investigators need to trust the result.
Network analysis.
The expensive schemes are rings, not individuals: a physician steering patients to a lab he secretly owns, a pharmacy filling scripts from a complicit prescriber. Graph models map the relationships across providers, patients, and facilities, and expose clusters that no claim-by-claim review would ever catch.
Document review.
Language models can read the clinical note and compare it against the billed code. That is where waste and abuse live: documentation that does not support the level of service charged.
The capability that actually changes the economics is prepayment scoring: running claims through these models before they are paid, so the suspicious ones are held or reviewed first. That turns the whole problem from recovery into prevention. Instead of chasing a fraudster after the fact, you never pay him at all.
Where the opportunity lies ahead
The winning startups here will not be the companies selling slightly better audits after the money is gone. They will be the companies that move detection into the payment pipeline, and, harder still, the ones that connect data across multiple payers. Cross-payer data will allow these companies to build continually improving agentic systems that detect fraud incrementally better every day.
The final step required here is mass adoption: getting risk-averse institutions like large payers to act on a probabilistic flag in a setting where a wrong accusation carries real consequences, and getting them to share data they have every incentive to hoard. That is a far easier problem than inventing the technology, and we have already done the hard part. The only open question is how long the healthcare industry will keep treating a hundred-billion-dollar leak as the cost of doing business.
About Kasra Khadem Kasra Khadem is a Partner at Pathlight Ventures, an early-stage venture capital firm in New York City, where he leads healthcare technology investing. Prior to joining Pathlight, Kasra was an Investor at Human Capital where he partnered with leading healthcare startups like Commure, Transcarent, and Ambience Healthcare.
Reader Interactions