Artificial intelligence has revealed a staggering gap in the detection of dangerous drinking habits, identifying nearly 60 times more at-risk patients than current standard screening methods in a large-scale primary care study.
Research from the University of Michigan, published in Drug and Alcohol Dependence, unleashed a natural language processing (NLP) tool on the electronic health records of over 133,000 patients. While standard diagnosis codes flagged just 820 individuals (0.62%) as having unhealthy alcohol use, the AI scanned clinical notes to identify an additional 47,514 patients (35.91%) showing signs of risky behaviour.
This massive discrepancy suggests that for every patient currently flagged by the system, dozens more may be flying under the radar, constituting a largely “hidden” population struggling with alcohol use who never receive a formal diagnosis.
“The findings suggest the majority of people drinking at risky levels are overlooked,” said Anne Fernandez, PhD, an addiction psychologist at the U-M Addiction Treatment Service. “Doctors can’t read every clinical note from every provider… but automated tools can do this quickly and easily”.
Crucial insights
The study offered crucial insights into who is being missed. The researchers recruited a subset of 170 participants to validate the findings using gold-standard self-report measures.
They found a stark difference in the clinical profiles of the two groups. Patients identified by standard methods (diagnostic codes) were significantly more likely to suffer from severe comorbidities, including higher rates of depression and anxiety, and were more likely to have already sought treatment.
In contrast, the patients identified solely by the AI represented a “previously unidentified population”. These individuals had similar rates of risky drinking but fewer severe symptoms of Alcohol Use Disorder (AUD) and lower rates of depression and anxiety.
This suggests the AI is capable of acting as an early warning system, spotting patients with “mild AUD or risky drinking” before their condition progresses to a more severe, clinically obvious stage.
Failure of standard screenings
The failure of standard screenings often stems from the limitations of structured data. Diagnosis codes and checkbox questionnaires rely on patients being forthcoming and clinicians having the time to administer screenings consistently.
However, crucial clues often exist in the free-text notes doctors type after a visit—subtle mentions of drinking habits that don’t make it into a formal diagnosis code. The NLP algorithm was designed to “read” these unstructured notes, using conditional logic to categorise text indicators related to consumption patterns and symptoms.
Despite the AI’s impressive reach, the researchers caution that it is not infallible. In the validation group, the AI identified 23 cases of risky alcohol use and 17 cases of AUD that standard screening missed, but the concordance with self-reported data was not perfect.
The authors conclude that NLP’s most significant potential lies in its ability to complement traditional screening. By combining the broad detection capabilities of AI with the precision of standard approaches, clinics could prioritise “overlooked” individuals for follow-up screening and intervention before their health deteriorates.
