The night shift was unusually calm. Mr. T, a 74-year-old admitted for pneumonia, had completed his evening antibiotics and was resting quietly. His blood pressure, heart rate, temperature, and oxygen saturation were all within normal limits. The nurse’s 2 a.m. note was brief but telling: “Less talkative tonight,” “Needed encouragement to walk to the bathroom,” “Left most of his dinner untouched.”

By morning, Mr. T was febrile, hypotensive, and confused. Within hours, he was in the ICU with septic shock.

Looking back, there had been a shift, but it was not in the numbers. It was in his behavior, his mobility, and his engagement with care. These were changes too subtle to set off any early warning system based on vitals alone. They were hiding in plain sight, in the notes we write every day.

The dangers of relying on vital signs only

Clinicians often rely on a combination of gut instinct and structured data such as vital signs, labs, and standardized scores as the key indicators of deterioration. Early warning systems like Modified Early Warning Scores (MEWS) and National Early Warning Score (NEWS2) are valuable, but they focus on quantifiable changes. They cannot easily capture “hesitant to get out of bed,” “slower movements during dressing changes,” or “requiring more assistance to eat.”

In busy hospitals, these cues get buried under piles of daily notes. Even the most attentive clinician can miss them, especially when caring for multiple patients or working long shifts. The tragedy is that many patients, like Mr. T, begin to show these signs hours to days before they crash.

Where natural language processing (NLP) fits in

NLP, a branch of artificial intelligence, can analyze free-text documentation at scale. Instead of skimming a few latest notes, NLP can read the entire chart in seconds, identifying repeated or escalating mentions of key phrases, even when spread across multiple clinicians’ entries.

For example, an NLP system could flag a patient whose last three progress notes mention:

• “Took longer to answer questions”
• “Needed repeated prompting during physical therapy session today.”
• “Reluctant to swallow pills, with more coughing with water”

Individually, each phrase might seem minor, but together, they could indicate early delirium, neurological decline, or aspiration risk long before vital signs shift.

Research supports the potential of NLP in this space. NLP models are capable of analyzing nursing notes to identify deterioration in inpatients hours to days earlier than vital signs alone. It is able to detect patterns of language describing functional changes correlating strongly with adverse outcomes, even in patients with stable labs and vitals.

Importantly, these cues often come from multiple sources including doctors, nurses, physical therapists, speech-language pathologists, and even the social worker. Humans rarely have the time or capacity to synthesize that much documentation in real time, but NLP can.

Implications for resource-limited settings

In high-resource hospitals, continuous telemetry, frequent lab testing, and advanced monitoring help detect decline. But in many parts of the world, these are not available. What is available in nearly every inpatient and outpatient setting is written notes only.

In such contexts, an NLP-based early warning system could transform care. It would require software to extract value from the data already being collected. First, a paper chart would need to be digitized and processed, and a basic EHR would be needed to run the algorithm in the background. This could make a difference where every ICU bed demands high-level resources that are already limited. In such settings, NLP and early intervention can avoid patient decline and high dependency care which, despite efforts, at times still lead to patient loss due to late intervention.

The importance of appropriate documentation

NLP functions as well as the information it receives. Algorithms can misinterpret context. To elaborate, if a nurse documents “patient less talkative,” it is important that she also explains the context. For instance, is it that his response was appropriate since it was 2 a.m. and it was his regular sleeping hours? Is it that his behavior was unusual because he would normally wake and be very talkative when being cared for at that time, or perhaps his behavior was unusual because it was 4 p.m. and he would usually be ambulating and interacting with his roommates?

Such thoughtful documentation, combined with other phrases such as “less steady,” “more withdrawn,” or “struggling to chew” over the past twenty-four hours could lead to more accurate and timely alerts. Inappropriate documentation could result in false alarms.

Another consideration is biases in training data which could also produce false alarms or missed detections, especially for populations underrepresented in the data. Transparency, rigorous validation, and clinician oversight are essential.

NLP should be seen as a complement to, not a replacement for bedside judgment. The most effective systems will integrate seamlessly into clinical workflows, surfacing meaningful alerts without adding to documentation burden or alarm fatigue.

Sometimes, the earliest signs of decline are not in the numbers but in the narrative across multiple entries during the course of admission. NLP offers a way to detect those signs more clearly, more consistently, and before they escalate into emergencies.

As clinicians, our words matter, not just for communication with colleagues, but as data that, when combined with technology, could save lives.

The challenge before us is to harness this potential responsibly, to build systems that amplify our vigilance, and to embrace tools that help us detect what our instincts already suspect. Most times the patient chart is speaking long before the vitals do, and the sooner we are able to “listen,” the more patients we can keep out of the ICU.

Jalene Jacob is a physician-entrepreneur.