AI in healthcare registration

Artificial Intelligence (AI) has become an integral part of the healthcare landscape. AI is regularly seen as the technology that will change everything: making work more efficient, reducing administration and improving care. At the same time, there is also a sense of realism. AI can do a lot and is improving rapidly, but we are not there yet. It is precisely in that tension between promise and practice that the real value lies. During the care event Zorg & ICT, Thomas Overmars and Ella Sijbesma discussed the role of AI in healthcare registration during their knowledge session.

Healthcare registration as a foundation

Thomas Overmars, Senior DBC Advisor at Zaans Medical Centre, and Ella Sijbesma, Business Area Manager Real World Monitoring at Performation, work with healthcare registration on a daily basis. During their knowledge session, they explain that healthcare registration forms the basis of almost everything within a care facility. It determines whether care provided is fully and correctly invoiced, whether processes run smoothly and whether adjustments can be made using the correct information. If that foundation is not in place, problems arise, including missed revenue, risks during checks and a distorted view of forecasts, capacity monitoring and care demand.

The limits of structured data

Traditionally, healthcare registration focused on structured data, such as DBCs, procedures and admissions. Checks were set up based on this data, with the aim of identifying anomalies. These include situations where a patient has been to the operating theatre, but no operation has been registered, or when an outpatient visit is registered at the same time as a nursing day. These kinds of alerts end up on a work list, after which administrative staff have to manually assess whether corrections are necessary. That process works, but it also has its limits. In some situations, more context is needed to determine whether something is correct or incorrect, and that context is often not found in structured fields but in free text.

The step towards unstructured data

That is why, in addition to the classic data checks, a second approach has been created using reporting such as doctor letters, nursing reports and lab reports. By scanning and analysing these texts, additional information becomes available that helps assess healthcare registrations.

This is where AI comes into the picture, not in the form of generative models such as ChatGPT, but as classification and forecasting models. These models label text and estimate how likely it is that a transaction has been correctly registered or missed.

Example of AI use in healthcare registration

A concrete example is wound treatment in the Emergency Department. Such a procedure may only be registered if specific conditions are met, such as the size of the wound, suturing and the use of anaesthetic. The AI model analyses reporting and searches for text fragments that support or contradict registration. Based on this, it calculates a probability score. These models are not yet perfect and can sometimes identify details that don’t immediately make sense to users. However, because users provide feedback on whether the alert was correct, the model is continuously learning.

From model to workplace

For healthcare facilities, the technology remains largely ‘under the bonnet’. What employees see is a worklist with alerts, with percentages indicating how likely it is that a procedure has been missed or incorrectly registered. This enables prioritisation. Instead of manually checking tens of thousands of consultations, the focus can be placed on the alerts with the highest likelihood. This saves significant time and increases the effectiveness of checks. In addition, the workload will be lower because less unnecessary checking is required.

The impact of this is significant. A missed peer consultation, for example, can quickly result in €150 in missed revenue and a missed chemotherapy session €1,200. For certain expensive drug registrations, the missed income is even higher. With AI, these types of losses are becoming more visible, without a commensurate increase in manual effort.

The role of samples

In addition to completeness, compliance also plays a crucial role. Audits by insurers are increasingly looking at process accountability rather than individual registrations. Nevertheless, sampling remains necessary. These samples often feel like roulette: you hope not to find anything, but if you do, a small error caused by extrapolation can have major financial consequences.

Compliance without roulette

AI offers an alternative here. By comprehensively assessing all consultations for the likelihood of compliance, a classification into different risk groups is created. Based on this, a practical division can be made into three categories:

• Rejection: high risk of non-compliance
• Assessment: cases of uncertainty
• Acceptable: sufficient confidence in compliance

The first two groups are actively monitored, while the final group is only randomly checked. To ensure support and transparency, this approach is only explicitly applied in consultation with health insurers. The result is reduced risk of high retrospective repayments, greater control over the process and a scalable approach to checks.

Practical lessons

Implementing AI isn’t plug and play. It requires thorough preparation. This includes considerations such as server capacity, computing power, new contracts and data processing agreements. Sharing patient-sensitive text data with external parties understandably raises questions for CISOs and privacy officers. This requires time, consultation and careful planning.

From a content perspective, introducing AI into healthcare registration also requires perseverance. The process involves testing, training, adjustment and re-testing, especially when there are multiple EPR versions or data flows within a healthcare organisation. AI sometimes reveals the very problems you didn’t know existed.

Additional tool

Perhaps the most important lesson is that AI is not a holy grail. It is important to consider which method of checking is appropriate in each situation. Sometimes this will be data analysis, text mining or AI, and in other cases, it will be a combination of these methods. In addition, AI will always produce an assessment. It does not know what it does not know, so blind faith is not an option. That is why it should be seen as an additional tool in the toolbox.

Looking ahead

The development of AI is not standing still. In addition to retrospective checks, there are opportunities to provide alerts at the point of registration. Think of EPRs that warn if something may be missing or incorrect in the healthcare registration, even before a registration is finalised. Further into the future, there are also opportunities for autonomous coding. This is the automatic derivation of procedures, DBCs and diagnoses based on patient record information. This already works well for straightforward care, while it remains challenging in complex clinical situations. However, the method continues to improve.

From checks to prevention

Healthcare registration is currently focused on checks, but the process may eventually shift from ‘registering and checking’ to ‘registering what is allowed’. Based on the reports, AI would then determine what can be registered, after which the doctor or nurse would assess this proposal. Upon approval, the transactions would be automatically registered. This could largely eliminate the need for retrospective checks. Whether we’ll ever really reach this point, and when, remains uncertain today. What is clear is that AI is already helping to make healthcare registration smarter, more efficient and more manageable. Not as a panacea, but as a powerful supplement. And that’s where the real value lies.