An Intelligence-Based Model for Identifying Advanced HF in Clinical Practice

An augmented intelligence-based model can identify patients with advanced heart failure and has a higher accuracy than clinician review.

A study published in JACC: Advances developed and incorporated an augmented intelligence-enabled workflow to identify patients with advanced heart failure (HF) into clinical practice.

Investigators from Northwestern Medicine in the United States sourced data for this study that was collected between 2007 and 2020 for the Heart Failure Registry, which aggregated patient records from 11 hospitals and more than 100 clinical sites. Predictive features of advanced HF were selected from 3 established HF risk scores. The predictive ensemble maximum likelihood model was evaluated for its ability to classify patients as stage D HF, stage C HF, and no HF by separating the dataset in an 80/20 split to form the training and test cohorts. The model was trialed in clinical practice between April 2021 and February 2022.

The patient population comprised 14,830 individuals with a mean age of 65.3 [SD, 16.6] years, 46.9% were women, they had a body mass index of 29.6 [SD, 7.8], 64.7% were White, they had a left ventricular ejection fraction of 54.5% [SD, 15.7%,] and 73.1% had hypertension. The patients had stage D HF (n=1312), stage C HF (n=6033), and no HF (n=7485).

The algorithm selected 3 clinical characteristics, 2 vital signs, 10 medications, and 13 diagnostic tests as the most robust predictors for advanced HF.

Endeavors such as this require a multidisciplinary team with experience in design thinking, informatics, quality improvement, operations, and health information technology, as well as dedicated resources to monitor and improve performance over time.

Among the test cohort (n=2992), the model had a positive predictive value (PPV) of 0.74 and sensitivity of 0.43 for predicting stage D HF, a PPV of 0.78 and sensitivity of 0.85 for predicting stage C HF, and a PPV of 0.88 and sensitivity of 0.89 for predicting no HF. Altogether, the model had an accuracy of 0.83.

The model outperformed a physician-reviewed subset of 100 patients with stages D or C HF, in which they had a PPV of 0.60 and sensitivity of 0.25 for predicting stage D HF and PPV of 0.79 and sensitivity of 0.91 for predicting stage C HF. The physicians had an overall accuracy of 0.75.

During the trial in clinical practice, 416 patients were classified by the model and a nurse coordinator after prospectively reviewing patient data. The model classified 294 patients as stage D HF and 122 as stage C HF. The coordinator agreed with the model in 50.3% of cases classified as stage D HF and 63.1% of cases classified as stage C HF.

These patients were recommended to receive a follow-up at a HF clinic (n=77), a review in 3 months (n=58), an evaluation in an advanced HF new access clinic (n=56), an evaluation in a general cardiology clinic (n=3), or no additional recommendations (n=207).

The major limitation of this algorithm is that incorporating this model into clinical practice does not remove the need for a significant amount of manual chart review for clinicians.

In this study, an augmented intelligence-based model was created to identify patients with advanced HF. The predictive model was found to have a higher accuracy than clinician review. Using this model in practice has the potential to help streamline patient identification such that patients are referred for appropriate, timely follow-up. “Endeavors such as this require a multidisciplinary team with experience in design thinking, informatics, quality improvement, operations, and health information technology, as well as dedicated resources to monitor and improve performance over time,” the study authors wrote.

Disclosure: An author declared affiliations with industry. Please refer to the original article for a full list of disclosures.

References:

Cheema B, Mutharasan RK, Sharma A, et al. Augmented intelligence to identify patients with advanced heart failure in an integrated health system. JACC Adv. Published online October 1, 2022. doi:10.1016/j.jacadv.2022.100123