Kidney Week

Abstract: FR-OR011

Machine Learning-Driven Digital Phenotyping of Accelerometer-Based Physical Activity Patterns and Mortality in Patients on Hemodialysis: The PROMOTE Study

Session Information

• Artificial Intelligence and Data Science Transforming Kidney Care: From Algorithms to Action
  November 07, 2025 | Location: Room 361A, Convention Center
  Abstract Time: 04:40 PM - 04:50 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Kadono, Hikaru, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Hikaru Kadono, MD

• Sakaguchi, Yusuke, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Yusuke Sakaguchi, MD, PhD

• Oka, Tatsufumi, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Tatsufumi Oka, MD, PhD

• Doi, Yohei, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Yohei Doi, MD, PhD

• Uwatoko, Ryuta, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Ryuta Uwatoko, MD

• Fukuda, Shungo, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Shungo Fukuda, MD

• Sugimachi, Ayaka, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Ayaka Sugimachi, MD

• Kawano, Yuki, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Yuki Kawano, MD

• Matsui, Isao, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Isao Matsui, MD, PhD

• Mizui, Masayuki, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Masayuki Mizui, MD, PhD

• Kaimori, Jun-Ya, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Jun-Ya Kaimori, MD

• Isaka, Yoshitaka, Osaka Daigaku, Suita, Osaka Prefecture, Japan

Yoshitaka Isaka, MD, PhD

Background

Sedentary behavior is prevalent in hemodialysis (HD) patients. While physical activity (PA) guidelines recommend 150–300 min/week of moderate-intensity PA (MPA) or 75–150 min/week of vigorous-intensity PA (VPA), we have previously identified low-intensity PA (LPA) as a more optimal form of PA for HD patients. However, such recommendations based on a single PA metric overlook the complexity of real-world PA patterns. Digital phenotyping of PA patterns unique to HD patients may offer more actionable recommendations.

Methods

The PROMOTE study enrolled 1,030 HD patients who were prospectively followed for three years. PA was measured using triaxial accelerometers at baseline and one year later. The cohort was randomly split into derivation and validation groups in a 6:4 ratio. High-dimensional PA data were reduced to two dimensions using Uniform Manifold Approximation and Projection (UMAP), followed by k-means clustering. Cox models were used to examine associations between PA patterns and all-cause death.

Results

Three clusters were identified from UMAP-assisted k-means clustering. A significant difference in the ratio of moderate-to-vigorous PA (MVPA) to total PA was found across clusters, with median [interquartile range] MVPA ratios of 3 [1–5] %, 4 [2–6] %, and 15 [10–15] %. The first two clusters were classified as the “LPA-dominant type”, and the third as the “MVPA-mixed type”. Clustering was replicated in the validation group. These phenotypes were consistent on both dialysis and non-dialysis days, and persisted after one year. The MVPA-mixed type exhibited better survival than the LPA-dominant type (HR 0.70, 95% CI, 0.49-0.99). The hazard of death declined steeply with increasing MVPA ratio, with the lowest hazard observed at 13-16%.

Conclusion

The MVPA-mixed pattern, identified as a digital phenotype of HD patients, was associated with better survival than the LPA-dominant pattern.

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025hy9dhgq8