Request a walkthrough
Research direction

Better decisions start
with better evidence.

A proposed study of evidence quality, uncertainty and nurse escalation in heart-failure care.

Proposed retrospective study · Not a current product feature
The patient-centered question

Can AI help care teams recognize who needs closer attention, even when the information from home is incomplete?

Our goal is more timely attention and better access to ongoing care. The intended setting is RN-supervised navigation, where recommendations must be understandable and the limits of the evidence must be visible.

Why the answer is unknown

Care data is messy.
Errors have consequences.

Combining data sources does not establish that an escalation recommendation is reliable, useful or safe.

Signals arrive unevenly.

Home readings may be sparse or noisy. Navigator notes are unstructured. Clinical records can describe different moments or disagree.

Labels are hard to establish.

Escalation events are uncommon, and deciding when nurse review was warranted requires independent clinical adjudication.

Workload changes the answer.

False negatives can harm patients. Excessive alerts and uncertain cases also consume limited nurse-review capacity.

Candidate method for investigation

Separate the patient’s risk
from the quality of the evidence.

The proposed direction is a model that represents both the clinical signal and the reliability of its inputs: when a reading was taken, where a note came from, and whether the sources disagree.

It would evaluate whether the evidence supports an escalation estimate or requires further review, with the supporting observations and uncertainty visible to the RN.

The technical question.

Can a joint model of escalation risk and evidence quality improve which cases reach nurse review when device participation, note quality and clinic workflows change?

Missing-data modeling and the option to decline a prediction already exist. The research must establish whether a specific method for handling conflicting, uneven care data adds value beyond those approaches at the same total nurse-review workload.

This is a candidate research direction. The architecture, technical contribution and feasibility remain to be established with the research team. Adding features alone would not establish a new method.

Related methods that inform the comparison

GRU-D: modeling missingness and time intervals
SelectiveNet: learning prediction and rejection together

Proposed initial experiment

A focused question.
A test that can fail.

Begin with heart failure and retrospective evaluation. Study whether explicit information about missing readings, timing gaps and disagreement between sources improves the selection of cases for nurse review.

Hypothesis

At a fixed total nurse-review workload, a model given explicit missingness, timing-gap and source-disagreement features will identify more clinician-adjudicated escalation events than the same model without those features.

Comparisons

Compare protocol rules, standard prediction methods, established missing-data and selective-prediction approaches, and a candidate model with evidence-quality components removed. The initial feature ablation tests signal value; it does not by itself establish methodological novelty.

Clinical labels

Define the escalation event and time window before evaluation. Use independent clinical adjudication and report disagreement.

Evaluation

Measure sensitivity, missed events, precision, calibration, abstention rate and coverage, including subgroup results where sample sizes support meaningful estimates. Count both alerts and abstentions that require RN attention in the total review burden.

Generalization

Use only information available at the prediction time, and hold out patients and later time periods. Test a separate clinic if suitable data and enough adjudicated events are available.

The hypothesis is not supported if evidence-quality signals add no meaningful benefit, or if the benefit disappears outside the development data.

Research status

Evidence before
clinical deployment.

The method and acceptance criteria are still being defined. Planning includes selection of a principal investigator and research team.

Dr. Fazila Lalani is Amana’s clinical advisor for the proposed disease-specific escalation direction, beginning with heart failure.

What the first study must establish.

  • Appropriate data access and research permissions
  • Enough adjudicated events for a meaningful test
  • A contribution beyond existing prediction methods
  • Useful uncertainty estimates across patient groups

Retrospective results would not establish safety or benefit in live care. Any future use would require further evaluation, clinician-approved boundaries and RN review and override.

Longer-term direction

From escalation
to the next care step.

The first study would test who needs RN review. Selecting the next care action is a later research question, followed by possible expansion to other programs and conditions. Condition-based screening, beginning with diabetes, is a separate roadmap concept.

Clinical outcomes would be evaluated independently of payment outcomes. Reimbursement is not a label for clinical correctness.

Discuss research collaboration

Grounded in an operating platform.

Today’s Amana platform supports care coordination, monitoring workflows and documentation. The intended research users are RN-supervised teams supporting people with heart failure. If validated, the work could help them decide which cases need closer review within the nursing capacity available.

The research asks what new evidence and methods are needed for reliable decision support. The proposed initial study focuses on escalation and uncertainty in heart-failure care.

See the platform today

Help investigate a consequential question.

We welcome conversations about clinical validation, evidence-quality modeling and careful evaluation.

Talk with Amana

Inside the Amana platform