In brief
An NIMH K23 proposal to build machine-learning models from electronic health record and claims data to estimate six-month severe mood-crisis risk in young people with bipolar disorder, and to co-design a clinical decision support tool. The models are planned, not yet validated.
What this article is about
Quick Answer
An NIMH K23 proposal to build machine-learning models from electronic health record and claims data to estimate six-month severe mood-crisis risk in young people with bipolar disorder, and to co-design a clinical decision support tool. The models are planned, not yet validated.
Student takeaways
Key Takeaways
- This is a National Institute of Mental Health K23 career development award (a funded proposal), so it describes aims and methods rather than results, and the risk models and decision support tool have not yet been built or validated.
- The abstract characterizes bipolar disorder as among the deadliest and costliest psychiatric disorders in young people, with recurrent depression and mania that disrupt functioning and substantially raise the risk of suicide and premature death.
- Aim 1 proposes using machine learning on electronic health record and claims data from more than 13,200 young patients (ages 15 to 39) to estimate six-month risk of severe mood crisis events, defined as mood-related inpatient hospitalizations or emergency visits.
- Aim 2 proposes engaging clinicians and patients through a modified Delphi approach and qualitative interviews to develop and evaluate a translatable clinical decision support tool for personalized prevention and early intervention.
- The project is explicitly foundational work meant to enable a future, larger study that would externally validate and rigorously evaluate the models and tool before clinical use.
Student summary
Why This Research Matters
This record is a funded research proposal (a mentored K23 career development award from the National Institute of Mental Health) rather than a completed study. That means it sets out aims, methods, and rationale, but does not report results, accuracy figures, or a finished tool. The project's goal is to build computer models that use electronic health record and insurance claims data to estimate the risk that a young person with bipolar disorder will have a severe mood crisis within a defined time window, and then to design a clinical decision support tool that could help guide care.
Bipolar disorder, often shortened to BD, is a serious mental illness marked by recurring mood episodes of depression and mania. The abstract describes it as among the deadliest and most costly psychiatric disorders in young people, because these episodes disrupt daily functioning, substantially increase the risk of suicide and early death, and often require emergency or inpatient care. An important clinical idea in the proposal is that each additional severe mood episode tends to worsen a person's long-term outlook, which is why preventing these episodes matters so much. A major barrier, the abstract argues, is that clinicians currently lack widely usable tools to identify which patients are at risk of a crisis within a specific interval, and to help tailor care to the individual.
To address this, the plan has two main aims. In the first aim, machine learning methods, a type of data analysis that finds patterns across large amounts of information, would be used to estimate the risk of a severe mood crisis event, defined as a mood-related inpatient hospitalization or emergency visit, over six-month intervals. This work would draw on rich, long-term health system data from two large learning healthcare systems, covering more than 13,200 young patients with bipolar disorder aged 15 to 39. In the second aim, the researcher would work directly with clinicians and patients, using a structured consensus method called a modified Delphi approach along with qualitative interviews, to develop and evaluate a clinical decision support tool. The intent is that the tool could be translated into real clinical settings and help guide personalized prevention and early intervention.
Because this is a career development award, part of its purpose is training the investigator, a doctorally prepared psychiatric nurse practitioner, in advanced predictive analytics, stakeholder-engaged intervention development, and health systems research. The abstract states that this work is meant to lay the foundation for a future, larger study that would externally validate and rigorously test the risk models and the decision support tool. In other words, the models described here would still need to be validated in new populations before anyone could trust them in practice.
For a nursing student, the appraisal lessons are significant. First, this is a plan, so it would be wrong to say the models work or are accurate; they have not been built and tested in the way that matters clinically. Second, models built from electronic health record and claims data reflect the quality and biases of that data. If certain groups are underrepresented or receive different care, a model can inherit and even amplify those inequities, which is a real cultural-safety and fairness concern. Third, a risk score is a support for clinical judgment, never a replacement for it; predictions describe probabilities across groups, not certainties about a specific person.
The topic also demands sensitivity. Bipolar disorder and its link to suicide are serious, and content like this should be handled in an educational, non-alarmist way. A prediction tool must never be used to label, stigmatize, or deny care to a patient, and any conversation about suicide risk belongs within a supportive, safety-focused, clinically appropriate framework. Nurses caring for young people with bipolar disorder can already apply the spirit of this work: watch for warning signs of mood escalation, support medication adherence and routines, build trusting therapeutic relationships, and know their facility's pathways for urgent mental health support. If a tool like this is eventually validated, it could help direct attention and resources, but it would work best as one input alongside careful, compassionate human assessment.
Source abstract
Study Overview
Project Summary Bipolar disorder (BD) is among the deadliest and most costly psychiatric disorders in young people due to its severe and recurrent mood episodes of depression and mania which disrupt functioning, substantially increase the risk for suicide and premature death, and frequently require emergency or inpatient care. As each subsequent mood episode worsens prognosis, prevention of severe mood events in young people with BD is central to mitigating its enormous personal and societal burden. However, prevention is hindered by the lack of widely deployable tools to identify which affected individuals are at risk of a severe mood crisis event within a specific interval and which can guide individualized care. Through this mentored K23 award, the candidate, a PhD-prepared psychiatric nurse practitioner, will build upon her background in early intervention for BD, data- driven analytic approaches, and qualitative methods. Her program of training and research are designed to leverage real-world data and advanced analytic machine learning methods to efficiently identify young individuals with BD at risk for severe mood events and develop a deployment-focused clinical decision support intervention in partnership with clinicians and patients that could be rapidly translated to clinical care (NIMH Strategic Objectives 4.1 and 4.2). Through planned training activities, the candidate will gain a strong skillset in advanced predictive analytics and machine learning using electronic health record (EHR) and administrative data, mixed methods for stakeholder engaged intervention development, embedded health systems research, and BD clinical epidemiology. She will leverage robust, longitudinal health system data from two learning healthcare systems in the Mental Health Research Network, HealthPartners and Kaiser Permanente Northern California, and engagement with clinicians and patients where care is delivered. In Aim 1, rigorous machine learning methods will be used to estimate risk of severe mood crisis events, as indicated by mood-related inpatient hospitalization or emergency visits, over six-month intervals based on rich longitudinal EHR and claims data in a large sample of over 13,200 young patients with BD aged 15-39 years. In Aim 2, to maximize the translational impact of the models, clinicians and patients will be engaged, using a modified Delphi approach and qualitative interviews, in development and evaluation of a clinical decision support tool to guide personalized prevention and early intervention for BD mood crises. This research is a critical step in the candidate's long-term goal of leveraging data-driven approaches to improve individualized, patient-centered delivery of mental health services for individuals in the early course of BD and other serious mental illnesses. Her clinical and research background, expert mentoring team, and embedded research environment ideally positions her to accomplish the research and training aims, building the foundation for a next-step R01 that will externally validate and rigorously evaluate the risk prediction models and decision support tool developed in this proposal.
Evidence appraisal
Main Findings
- This is a National Institute of Mental Health K23 career development award (a funded proposal), so it describes aims and methods rather than results, and the risk models and decision support tool have not yet been built or validated.
- The abstract characterizes bipolar disorder as among the deadliest and costliest psychiatric disorders in young people, with recurrent depression and mania that disrupt functioning and substantially raise the risk of suicide and premature death.
- Aim 1 proposes using machine learning on electronic health record and claims data from more than 13,200 young patients (ages 15 to 39) to estimate six-month risk of severe mood crisis events, defined as mood-related inpatient hospitalizations or emergency visits.
- Aim 2 proposes engaging clinicians and patients through a modified Delphi approach and qualitative interviews to develop and evaluate a translatable clinical decision support tool for personalized prevention and early intervention.
- The project is explicitly foundational work meant to enable a future, larger study that would externally validate and rigorously evaluate the models and tool before clinical use.
Practice transfer
Clinical Relevance
- Nurses caring for young people with bipolar disorder should watch for warning signs of mood escalation and support routines, medication adherence, and strong therapeutic relationships.
- Because each severe mood episode can worsen long-term prognosis, early recognition and timely intervention are clinically important goals that this line of research seeks to support.
- A risk score, if one is eventually validated, should inform and support clinical judgment, never replace it, since predictions describe group probabilities rather than certainties about an individual.
- Models built from electronic health record and claims data can inherit biases in that data, so nurses and teams must guard against tools that could worsen inequities for underrepresented groups.
- Risk information about a serious condition must be handled sensitively and safely, and never used to label, stigmatize, ration, or deny care; suicide-related concerns require supportive, safety-focused clinical pathways.
Faculty notes
Educational Relevance
This National Institute of Mental Health K23 proposal is a strong vehicle for teaching critical appraisal of predictive analytics in mental health nursing. It is a funded career development award, not a completed study, so students should recognize that the risk models and clinical decision support tool are planned, not validated. Aim 1 would use machine learning on electronic health record and claims data from over 13,200 young patients (ages 15 to 39) with bipolar disorder to estimate six-month risk of severe mood crisis events (mood-related hospitalization or emergency visits). Aim 2 would engage clinicians and patients through a modified Delphi process and qualitative interviews to build a deployable decision support tool. Use this to discuss algorithmic bias and health equity (models inherit the biases of their training data), the difference between prediction and causation, external validation as a prerequisite for clinical use, and the principle that risk scores augment rather than replace clinical judgment. It also raises safety and stigma concerns: predictions about suicide-related risk must never be used to label, ration, or deny care. Encourage students to weigh the promise of earlier, personalized intervention against the ethical and practical demands of implementing algorithms in psychiatric care.
Critical appraisal
Limitations
- This is a funded K23 proposal, not a completed study; no models, accuracy figures, or outcomes are reported, and the tools described do not yet exist in validated form.
- The planned models would rely on electronic health record and claims data, which reflect the quality, completeness, and biases of routine clinical documentation and billing.
- The work is explicitly preliminary, with external validation and rigorous evaluation reserved for a future, larger study, so clinical usefulness cannot yet be judged.
Classroom use
Discussion Questions
- Why is it important to recognize that the risk models in this proposal are planned rather than validated and ready for use?
- What is bipolar disorder, and why does the abstract describe it as among the deadliest and most costly psychiatric disorders in young people?
- Why might preventing each severe mood episode matter for a young person's long-term outlook?
- How is a 'severe mood crisis event' defined in this proposal, and what are the strengths and limits of that definition?
- What is machine learning, and what are the risks of building clinical prediction models from electronic health record and claims data?
- How can algorithmic bias in a risk model create fairness and cultural-safety problems for certain groups of patients?
- Why is external validation necessary before a prediction model is used in real clinical care?
- How should a risk score relate to a nurse's or clinician's own judgment when caring for a patient?
- What ethical concerns arise from predicting risk related to suicide, and how can care teams handle such information safely and without stigma?
- What can nurses do now to support young people with bipolar disorder, even before a tool like this exists?
Search-ready answers
Frequently asked questions
Does this project provide a working tool to predict mood crises?
No. It is a funded proposal. The risk models and decision support tool are planned and would need to be built and validated before clinical use.
What is bipolar disorder?
A serious mental illness involving recurring episodes of depression and mania that can disrupt functioning and raise the risk of suicide and premature death.
What would the models try to predict?
The six-month risk of a severe mood crisis event, defined as a mood-related inpatient hospitalization or emergency visit.
What data would be used?
Electronic health record and insurance claims data from more than 13,200 young patients in two large healthcare systems.
Why involve clinicians and patients in Aim 2?
To make the decision support tool practical and translatable to real care, using a modified Delphi process and interviews.
Are these predictions certain for an individual patient?
No. Predictions describe probabilities across groups, so a score should support, not replace, careful clinical judgment.
Could such a model be biased?
Yes. Models built from routine health data can inherit the data's biases and worsen inequities, which is why fairness and validation matter.
Why must the models be externally validated?
To confirm they perform well in new, independent populations before anyone relies on them clinically.
Could a risk score be used to deny someone care?
It should never be. Risk information must not be used to label, stigmatize, ration, or deny care, especially for a serious condition.
What can nurses do now?
Support young people with bipolar disorder through vigilance for mood escalation, help with adherence and routines, trusting relationships, and knowing urgent mental health pathways.