For two decades, the United States has watched suicide rates climb in a steady, devastating arc. But for American Indian and Alaska Native (AIAN) individuals, this trend is not just a statistic—it is a crisis of disproportionate intensity. In 2023, while the national age-adjusted suicide rate stood at 14.1 per 100,000, the rate for AIAN populations soared to 23.8 per 100,000, marking a 44.2% increase since 2011.
This disparity is the result of a complex, painful confluence of intergenerational trauma, historical marginalization and a chronic lack of mental health resources. In the high-pressure environment of an emergency department (ED), where patients are often at their most vulnerable, the ability to identify someone at risk of suicide in a matter of minutes can be the difference between life and death. Yet, the tools currently used to do this are often failing the people who need them most.
A new study published in the Journal of Medical Internet Research (JMIR) suggests that the answer isn’t to replace human clinicians or traditional screening, but to augment them with machine learning. By analyzing electronic health records (EHR) in partnership with the White Mountain Apache Tribe and the Indian Health Service (IHS), researchers found that AI can act as a critical “safety net,” catching high-risk patients who slip through the cracks of standard medical questionnaires.
The Limits of the Standard Checklist
Currently, many hospitals rely on the Ask Suicide-Screening Questions (ASQ), a brief, four-item tool developed by the National Institute of Mental Health. It is designed to be swift and non-invasive, asking patients about their desire to die or previous attempts. In theory, it is a vital first line of defense. In practice, the study found it was significantly underperforming in AIAN populations.
The research revealed a sobering reality: the ASQ had a sensitivity of only 17.8%. This means it identified fewer than one in five patients who subsequently attempted or died by suicide within 90 days. Even more alarming, no one who died by suicide in the study group had screened positive on the ASQ during their ED visit.
As a former software engineer, I see this as a classic “edge case” problem. Tools developed for majority populations are often applied to minority populations without proper validation. When a tool doesn’t account for cultural differences in communication or the specific nature of trauma within a community, it doesn’t just fail—it creates a dangerous illusion of safety.
AI as a Clinical Safety Net
To bridge this gap, researchers developed a machine learning model tailored specifically for the IHS setting. Unlike a questionnaire, which captures a snapshot of a patient’s mood in a single moment, the ML model looks at the “long game.” It analyzes longitudinal EHR data, including demographics, clinical diagnoses, medication history, and patterns of healthcare use over the previous five years.
The results were stark. When set to a broader risk threshold, the ML model achieved a sensitivity of 78.2%—vastly outperforming the ASQ. However, the researchers weren’t looking to replace the human element. Instead, they tested two ways of integrating AI into the clinical workflow: parallel testing and serial testing.
- Parallel Testing: The patient is flagged if either the ASQ is positive or the AI identifies them as high risk. This is a “wide net” approach.
- Serial Testing: The AI acts as a filter; only those flagged by the model are then given the ASQ. This is a “precision” approach designed to reduce false alarms.
The study found that parallel testing was the most effective pragmatic strategy. By using AI to verify negative screens or flag patients when a screen was missed—which happened in 37.3% of visits—the sensitivity jumped to 79.5%. It preserved the universal nature of screening while ensuring that the AI caught those the checklist missed.
Comparing Risk Identification Strategies
| Strategy | Sensitivity (Detection Rate) | PPV (Accuracy of Flags) | Clinical Role |
|---|---|---|---|
| ASQ Screening Alone | 17.8% | 2.2% | Baseline Tool |
| ML Model Alone | 78.2% | 1.2% | Data-Driven Insight |
| Parallel (ASQ + ML) | 79.5% | 1.2% | Safety Net |
| Serial (ML then ASQ) | 12.9% – 16.4% | 2.9% – 5.0% | Precision Filter |
The Ethics of the False Positive
One of the most contentious points in AI healthcare is the “false positive”—when the algorithm flags a patient as high risk, but they are not actually in danger. Critics argue that too many false alarms lead to “alert fatigue” for doctors and unnecessary interventions for patients.
The researchers in this study offer a compelling counter-argument: the cost of a false negative (missing a suicide attempt) is infinitely higher than the cost of a false positive. They compared the AI’s accuracy to mammography, which often has similarly low positive predictive values (PPV) but is accepted because the benefit of early detection outweighs the burden of a follow-up biopsy.
a “false positive” in a mental health context isn’t necessarily a waste of time. A patient flagged by AI might not be suicidal, but the algorithm may have picked up on other underlying conditions—such as substance use or chronic pain—that a clinician can then address, providing a gateway to care that otherwise would have been ignored.
The Path Forward
The study concludes that while AI is a powerful tool, it cannot stand alone. It depends on historical data, meaning a patient who is new to the health system remains “invisible” to the algorithm. This is why the human-led ASQ screen remains essential; it captures the now, while the AI captures the history.
The next critical step for the Native-RISE project and similar initiatives is to move from retrospective analysis to prospective clinical trials. The goal is to determine if these “safety net” alerts actually lead to lower suicide rates in real-time clinical settings.
Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always seek the advice of your physician or other qualified health provider with any questions you may have regarding a medical condition.
If you or someone you know is in crisis, please contact the 988 Suicide & Crisis Lifeline by calling or texting 988 in the US and Canada, or calling 111 in the UK. These services are free, confidential, and available 24/7.
We want to hear from you. Do you believe AI should be used as a “safety net” in emergency medicine, or are the risks of false positives too high? Share your thoughts in the comments below.
