Decision Support Interventions
Predictive Decision Support Interventions: Risk Management Summary
Last reviewed September 30, 2026 · Reviewed every quarter
How Careexpand manages the risks of the predictive decision support interventions it supplies in Careexpand EHR v2024.11.1, under 45 CFR 170.315(b)(11).
| Health IT Module | Careexpand EHR v2024.11.1 |
| Applies to | 45 CFR 170.315(b)(11), Decision Support Interventions |
| Last reviewed | September 30, 2026 |
| Reviewed | Every quarter, and whenever an intervention changes |
| Contact | Careexpand LLC, 9400 North Central Expressway, Dallas, Texas 75231. support@careexpand.com |
What this covers
Careexpand EHR includes four features that use a large language model to support clinicians. Under the ONC Health IT Certification Program they are Predictive Decision Support Interventions supplied by Careexpand:
- Protocol diagnosis insights. When a clinician sets up a remote-monitoring care protocol, it suggests symptoms to monitor, questions to ask, and a medication checklist for the diagnosis entered. It uses the diagnosis title only, not patient data.
- Care protocol suggestion from an encounter. It suggests a remote-monitoring protocol, including its schedule and what to monitor, from the visit documentation.
- CareCheck assessment. It summarizes automated monitoring calls. It covers medication adherence, symptoms and their severity, and whether the call content supports each diagnosis. It also produces follow-up actions.
- Previsit triage. From an automated call before an appointment, it flags patients who may need attention. It also proposes documentation entries for the clinician to accept or reject.
Each feature runs a third-party general-purpose model served by Groq, Inc. Careexpand does not train or fine-tune these models and does not use customer data to train any model. Funding: Careexpand's own funds; no external or government funding.
How we analyze risk
We assess each feature against the eight characteristics ONC names: validity, reliability, robustness, fairness, intelligibility, safety, security and privacy. For every feature we record what it receives, what it returns, who acts on the output, how it could fail, and the potential consequence for a patient.
The main risks we identified:
- output that is clinically inappropriate for a particular patient;
- output that differs between identical requests;
- errors carried over from call transcription, which may affect some speakers more than others;
- clinicians relying on a suggestion without reviewing it.
How we mitigate risk
A clinician stays in the loop
- Protocol suggestions only pre-fill a form. Nothing is saved until the clinician reviews and submits it.
- Previsit documentation suggestions stay pending until a clinician accepts or rejects each one.
- CareCheck creates its follow-up actions as tasks for the care team to act on.
Clear labeling
- Every output is marked "AI-generated · Predictive DSI" wherever it appears.
- Each output links to the feature's source attributes: its intended use, its limitations, and the other information ONC requires.
- Clinicians can record Accept or Dismiss, with a comment, on each output.
Constrained output
- Each feature returns structured data in fixed categories.
- Values outside those categories are rejected or normalized.
Organization control
- Organizations can turn protocol diagnosis insights and encounter protocol suggestions on or off, and hide them from specific roles.
- CareCheck runs only for patients enrolled in a care protocol. Previsit triage runs only when the organization uses previsit calls.
Transparency about evidence
- These features have not yet been evaluated quantitatively, and their source attributes say so.
Security and privacy
- Only signed-in users can request a protocol diagnosis insight or an encounter protocol suggestion, and only if their role grants the matching permission.
- Only users whose role grants decision-support administration can change the settings, the source attributes, or export feedback.
- Requests to the model provider travel over an encrypted (HTTPS) connection. The credential for the provider is kept in the platform's secret configuration, not in the software.
- Careexpand EHR does not keep a separate log of the text sent to the model or of what it returns. Outputs are kept only where the clinician uses them, in the patient's record.
- Each feature receives only the data it needs for the request (see Governance below).
- A structured output format and clinician review limit the effect of misleading text inside a transcript or note.
How we monitor the interventions
Every quarter, and whenever a feature's model, instructions or output format changes, Careexpand's clinical team reviews the feature. The review covers the Accept and Dismiss feedback clinicians recorded and a sample of the feature's outputs, checked for clinically incorrect or unsupported content. The sample depends on the feature:
- Stored outputs. CareCheck assessment and previsit triage store their outputs, so the review samples those from the quarter.
- Generated for the review. Protocol diagnosis insights and encounter suggestions do not store their outputs, so the review generates them: diagnosis insights for a fixed set of diagnosis titles, and encounter suggestions for a sample of the quarter's encounters.
For fairness, the same review compares outputs, and feedback where it is linked to a patient, by the patient's preferred language and age band where the record holds them. It looks for groups whose output is dismissed more often or is of lower quality. Protocol diagnosis insights receive no patient data, so their review instead checks outputs across diagnosis titles in different languages and wording, and for conditions specific to sex or age.
No quantitative performance or fairness threshold has been set yet. Findings are recorded and, where needed, lead to changes to the feature and to its source attributes.
Governance: how data is acquired, managed and used
What each feature uses
Only what it needs for one request:
- the diagnosis title;
- the visit documentation;
- the transcript of an automated call made to a patient enrolled in the relevant service.
Where it goes
The data is sent to the model provider only to produce that request's output. Careexpand has set the provider not to retain request or response data (zero data retention).
What is kept
CareCheck and previsit outputs are stored in the patient's record in Careexpand EHR, where clinicians review them. Protocol suggestions are not stored; only what the clinician submits is saved.
When a feature changes
Any change to a feature's model, instructions or output format is reviewed by a clinical subject-matter expert. The same release updates that feature's source attributes.
Questions
Careexpand LLC, support@careexpand.com