Why Medical Record Review Needs To Move Beyond The Rearview Mirror

Arpan Saxena is the COO/CIO at basys.ai (based out of Harvard University), a leading healthcare AI solutions company.

gettyHealthcare has become remarkably good at documenting what happened to a patient. The harder problem is recognizing what those records are telling us before something goes wrong.

Medical record review (MRR) remains one of healthcare’s most important quality safeguards. It helps identify gaps in care, assess compliance and give regulators and healthcare organizations a clearer picture of a patient’s journey. But much of that value arrives retrospectively.

A record is reviewed. A problem is identified. A surveyor finds a deficiency. A quality team investigates. Corrective action follows.

That model is necessary but increasingly insufficient. The next evolution of MRR should help organizations recognize emerging quality risks early enough to change the future.​

A single adverse outcome may be preceded by medication changes, repeated symptoms, missed follow-ups, abnormal laboratory values, nursing observations and subtle shifts in a patient’s condition. Individually, none may indicate a quality failure. Together, they may tell a very different story.

The challenge is that healthcare information rarely arrives as a clean narrative. Relevant evidence may be scattered across months of clinical documentation and multiple systems, providers and formats.

Traditional review asks a human to reconstruct that story, often after an event has occurred. AI creates an opportunity to change the timing.

Instead of asking only, “What happened?” medical record review can increasingly ask, “What is beginning to happen, and where should someone look more closely?”

There is a temptation to treat MRR primarily as a summarization problem: Give an AI system hundreds of pages and ask it to identify what matters. Healthcare quality is more complicated.

A clinically important review involves distinct tasks. Did the system retrieve the right evidence? Did it understand the clinical context? Did it correctly interpret the applicable requirement? Did it distinguish a genuine concern from normal clinical variation? Can a reviewer trace the conclusion to the source record?

A system could perform well on four tasks and poorly on the fifth. A single aggregate “accuracy” score may hide the difference. Healthcare should, therefore, evaluate AI-assisted MRR at the level of the decisions it makes.

We already accept this principle elsewhere in healthcare. Complex clinical or payment decisions are not adequately evaluated by asking whether the technology was simply “accurate.” We measure the individual steps that determine whether the final conclusion can be trusted.

This becomes particularly important as AI moves from retrospective review toward proactive quality monitoring. Finding more anomalies is not necessarily the goal.

Imagine a system that flags hundreds of concerns but cannot distinguish meaningful deterioration from routine variation. Detection may increase, but clinicians and quality teams inherit an expanding queue of low-value alerts.

Every unnecessary review consumes attention that could have been directed toward a patient or facility with genuine risk. At scale, poorly calibrated AI could reproduce a familiar healthcare problem: alert fatigue.

The objective should not be maximum detection. It should be better signal. That requires systems designed around clinical reasoning, uncertainty and human review rather than systems optimized simply to produce an answer.

This is where healthcare oversight can begin to change.

CMS has emphasized using medical record information to help providers address issues before they become survey deficiencies or negatively affect patients, while equipping surveyors with richer information before they arrive on-site. That points toward something larger than automating today’s review process.

Imagine MRR operating as a continuous intelligence layer.

Records could be evaluated against defined quality requirements as information becomes available. Potential concerns could be surfaced alongside clinical evidence. Patterns across patients or time could help distinguish an isolated documentation issue from a systemic quality risk.

Providers could investigate issues earlier. Surveyors could arrive with a clearer understanding of where attention may be most valuable. Limited oversight resources could be directed toward areas showing stronger signals of risk.

The survey would remain important. Human judgment would remain essential. But both would start with better information.​

Making this transition responsibly requires a different approach to healthcare AI.

Rather than asking one model to read a record and declare whether there is a problem, the process can be separated into checkpoints: evidence retrieval, clinical interpretation, requirement mapping, risk identification and human validation. Each checkpoint can be evaluated independently.

When something goes wrong, organizations can identify where it went wrong. When requirements change, the affected component can be updated and revalidated. When the system is uncertain, the case can be escalated rather than forcing a binary conclusion.

This architecture may sound less exciting than fully autonomous AI. In healthcare, that is precisely the point.

The goal should not be removing humans from MRR. It should be giving them a timely and interpretable view of information that no individual reviewer could continuously synthesize at scale.​

For decades, medical record review has helped healthcare look backward and determine whether appropriate care occurred. AI gives us an opportunity to move that window forward.

The most important measure of progress will not be how quickly a system summarizes a thousand-page record or how many potential deficiencies it flags. It will be whether the right signal reaches the right person early enough to matter.

That is the transition healthcare quality needs: from documenting problems to recognizing patterns, from periodic inspection to continuous intelligence and, ultimately, from learning what went wrong to having a better chance of preventing it.​

This article was co-written with CEO and co-founder Amber Nigam, a Forbes Business Council member.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?