The Anatomy of an Alert: When AI Sees Illness Before Symptoms Appear

Algorithms can forecast which patients are likely to return to the hospital. Whether anyone has the resources to intervene is another question.

The changes were subtle.

A laboratory value drifted. Another followed. A recent hospitalization joined years of clinical history already stored in the electronic health record. Alone, none of it seemed remarkable. Together, the data formed a pattern that only a machine learning model could recognize.

Long before she or her family saw herself getting sicker, the model estimated that she was at high risk of returning to the hospital.

A prediction arrived first. 

The question of how medicine should treat the person with that prediction is a far more difficult question.

You may have heard someone you know say, I wish someone had noticed sooner. If someone sees what is coming, perhaps there is still time to change it.

That question sits at the center of Lucas Zier's work.

A physician and researcher at UCSF, Dr. Zier studies what happens after a prediction has been made. Identifying patients at high risk for hospital readmission is, in many ways, the easier accomplishment. The harder problem begins once that prediction appears on a clinician's screen.

"We can identify who is at risk for readmission," he told me, "but then can we put them on a referral pathway to a specialty clinic.”

Prediction is the beginning, and knowing the future matters if it changes the present.

Machine learning has become effective at recognizing patterns across highly dimensional clinical data. It can estimate which patients are likely to be readmitted, whose conditions may worsen, or who may benefit from closer follow-up. However, the clinical usefulness lies in reducing illness for the patient.

The intervention exists outside the algorithm. It depends on the clinician who receives the alert, the nurse who has time to call the patient after discharge, the social worker who can arrange transportation or connect someone with community resources, and the healthcare system willing to invest in those services.

Every prediction creates a decision. If an algorithm warns that a patient is likely to deteriorate, is there someone prepared to act? If there is not, what has the prediction accomplished?

One additional problem is that clinicians work amid a constant stream of electronic alerts, some of which are not useful and ignored. Another notification can disappear into the background noise. 

Zier also notes that “When the right information is given at the right time it can help clinicians trust the alert more.” 

Lowering the patient detection threshold could detect more patients who could benefit from additional treatment but it also increases false alarms that clinicians see. Raising it reduces interruptions while allowing some patients to pass unnoticed. 

These are not simply technical adjustments. They determine which patients receive additional attention, and which do not.

For Jennifer James, a bioethicist at UCSF, those decisions begin with a question at a different layer of the algorithmic puzzle: whose experiences shape the problems medicine chooses to solve?

Her perspective suggests that the gap between detection, prediction and action is not only about whether someone responds to an alert; it is also about whether the detection algorithms are trained to spot illness across an array of clinical, demographic and socioeconomic situations. 

An algorithm can predict who is at risk of getting sick, but only if the training dataset includes every demographic person the health system plans to provide healthcare to. Additionally, clinical predictions are only useful if they address what patients actually need. 

That perspective changes the conversation about predictive AI. Accuracy remains essential, but it is no longer sufficient. Together, their work focuses on building a model that correctly identifies patients at risk, and improves care systems surrounding their disease.

Predictive medicine is changing what healthcare can know. The more difficult challenge is deciding what healthcare owes patients once that knowledge exists.

The future of clinical AI may depend less on building models that predict farther into the future but, more so building healthcare systems capable of responding to what those predications reveal.