Early Detection of Infection Using the Fitbit in Pediatric Surgical Patients (i-DETECT)

Ann & Robert H Lurie Children's Hospital of Chicago

Status: Enrolling

Conditions: Appendicitis; Appendicitis Acute; Appendectomy

Treatments: Device: Infection-Prediction Algorithm

Study type: Interventional

Funder types: Other

Identifiers

NCT06395636

IRB 2024-6701

Details and patient eligibility

About

The purpose of this study is to analyze Fitbit data to predict infection after surgery for complicated appendicitis and the effect this prediction has on clinician decision making.

Full description

We propose to investigate the use of objective near-real time data from the Fitbit consumer wearable device (CWD) for early detection of postoperative infection in children after appendectomy for complicated appendicitis, and its influence on clinician decision-making, time to first contact with the healthcare system, and postoperative healthcare use. SSI is usually associated with increased heart rate (HR) and reduced physical activity (PA), and sleep disturbances due to discomfort, pain, and fever.13-15 To help monitor patients post-discharge, CDWs can be used to detect physiologic changes, prompting early management.16,17 CWDs generate continuous, valid HR data comparable to clinical-grade HR monitor data for children, as well as objective PA and sleep data, which are good indicators of recovery.18-21 CWDs, then transmit these data in near-real time to a cloud-based system potentially accessible to a clinician. Although health systems have incorporated CWD data into electronic health records,16,17 use in post-discharge monitoring of pediatric surgery patients has been limited18 since it is difficult to monitor and interpret the large volumes of data generated by CWD in clinically meaningful ways.18 Machine learning (ML) methods, which reduce CWD data into clinically meaningful signals are needed.22 Since these algorithms are based on data from multiple CWD sensors, they are more accurate than threshold-based alerts.

We collected Fitbit data on 160 pediatric appendectomy patients21,23,24 and showed slower normative recovery PA trajectories in children with complicated versus simple appendicitis, and deviations from normative PA trajectory (decreased PA) before parents sought healthcare for complications.20 We then applied ML methods to Fitbit data of 80 post appendectomy patients with complicated appendicitis to predict infection. The preliminary algorithm predicted 90% of infections, 2 days before parental report. In parallel, we developed a proof-of-concept dashboard that delivers Fitbit data daily and on-demand in near real-time to clinicians. Using the dashboard, clinicians evaluated hypothetical post-discharge pediatric appendectomy scenarios with and without Fitbit dashboard data. Availability of Fitbit data (even without ML) substantially changed clinicians' likelihood of recommending ED care. While our early results are promising, a larger study is needed to definitively elucidate the association of changes in Fitbit data with postoperative infection and to assess the effect of Fitbit data on clinician decision-making and healthcare use. We propose to develop a ML algorithm for postoperative infection using Fitbit data of children 3-18 years old undergoing a appendectomy for complicated appendicitis at the Ann and Robert H. Lurie Children's Hospital of Chicago (LCH), a tertiary care children's hospital and two affiliated hospitals Loyola University Medical Center, a university hospital), and Central DuPage Hospital (CDH), a community hospital. Our two aims are:

Aim 1: Develop and externally validate an ML algorithm for postoperative infection. In addition to the 80 patients already recruited in our preliminary study, we will prospectively recruit 170 patients for a total of 250 from LCH for development and internal validation. We will then externally validate our infection ML algorithm using data on 122 appendectomy patients from LCH and its two affiliates.

Aim 2: Conduct a pre-post study to determine the effect of near real-time availability of the infection alert from Fitbit on clinical decision-making, time to first contact with the healthcare system, and healthcare utilization. We will place a Fitbit on 94 children after appendectomy recruited from LCH and its two affiliates, and send their surgeons daily reports of their recovery progress and near real-time, ML-based, clinical alerts of infection. In Aim 2a, we will use critical incident technique to qualitatively assess surgeons' decision-making after receiving Fitbit alerts and daily reports. In Aim 2b, we will compare median time to first contact with the healthcare system, healthcare use patterns (e.g., ED visits) and costs pre and post receiving alerts and daily reports.

Impact: This study is well aligned with NINR's priority to advance symptoms science. Developing CWD alerts to detect infection and evaluating their effect on clinical care have the potential to transform pediatric surgical care and pave the way for wide uptake of CWD. By proactively reaching to patients, this technology also has the potential to reduce existing disparities in seeking care.

Enrollment: 500 estimated patients

Sex: All

Ages: 3 to 18 years old

Volunteers: No Healthy Volunteers

Inclusion criteria

• children aged 3-18 years
• must be post-surgical laparoscopic appendectomy for complicated appendicitis (Appendicitis is categorized as complicated if perforation, phlegmon, or abscess was present at surgery.)

Exclusion criteria

• children who are non-ambulatory or have any pre-existing mobility limitations
• children who have a doctor-ordered physical activity limit >48 hours post-surgery
• children who have a comorbidity which will impact a patient's recovery
• children and/or parents who do not speak English or Spanish (Translation services beyond Spanish will not be available at this time)

Trial design

Primary purpose: Diagnostic

Allocation: Non-Randomized

Interventional model: Sequential Assignment

Masking: None (Open label)

---

500 participants in 2 patient groups

Aim 1 - Validation

No Intervention group

Description:

1a. Development and Internal validation * analyze Fitbit data (PA, HR, sleep) by applying ML methods to create an infection algorithm indicating onset of infection. 1b. External Validation * Once the ML classifier has been internally validated (using Lurie Children's data only) for its ability to detect the presence or absence of postoperative infection using LOSO cross-validation, where each subject is iteratively held out from the training data and used as a test set. External validation will involve applying this classifier to a newer cohort at LCH and cohorts at Loyola University Hospital and CDH and evaluating its performance.

Aim 2 - Implementation of Algorithm

Experimental group

Description:

2a. Exploratory \& Inductive analysis * one transcript will be coded to generate initial themes, using qualitative analytic software 2b. Time to first contact with the healthcare system \& Healthcare use * Cox regression model will be used to model the time to first contact, adjusted for covariates * All comparisons between the two groups will be tested using a chi-square test. Cost will be modeled as a continuous variable and is expected to be skewed, as is typical of cost data. We will use a general linear model (GLM) to model cost outcomes.

Treatment:

Device: Infection-Prediction Algorithm