JMIR Publications

ABSTRACT

Background:

ViSi Mobile has the capability of monitoring a patient’s posture continuously during hospitalization. Analysis of ViSi telemetry data enables researchers and healthcare providers to quantify an individual patient’s movement and investigate collective patterns of many patients. Erroneous values can exist in routinely collected ViSi telemetry data. Data must be scrutinized to remove erroneous records before statistical analysis.

Objective:

The objectives were to (1) develop a data cleaning procedure for a one-year inpatient ViSi posture dataset, (2) consolidate posture codes into categories, (3) derive concise summary statistics from the continuous monitoring data, and (4) study types of patient posture habits using summary statistics of posture duration and transition frequency.

Methods:

This study examined the 2019 inpatient ViSi posture records from Atrium Health Wake Forest Baptist Medical Center. First, two types of errors, record overlap and time inconsistency, were identified. An automated procedure was designed to search all records for these errors. A data cleaning procedure removed erroneous records. Second, data preprocessing was conducted. Each patient’s categorical time series was simplified by consolidating the 185 ViSi codes into 5 categories (Lying, Reclined, Upright, Unknown, User-Defined). A majority vote process was applied to remove bursts of short duration. Third, statistical analysis was conducted. For each patient, summary statistics were generated to measure average time duration of each posture and rate of posture transitions during whole day and separately during daytime and nighttime. A k-means clustering analysis was performed to divide the patients into subgroups objectively.

Results:

The analysis used a sample of 690 patients, with a median of 3 days of extensive ViSi monitoring per patient. The median of posture durations was 10.2 hours/day for Lying, 8.0 hours/day for Reclined, and 2.5 hours/day for Upright. Lying had similar percentages of patients in low and high durations. Reclined showed a decrease in patients for higher durations. Upright had its peak at 0 – 1 hours, with a decrease for higher durations. Scatterplots showed that patients could be divided into several subgroups with different posture habits. This was reinforced by the k-means analysis, which identified an active subgroup and two sedentary ones with different resting styles.

Conclusions:

Using a massive ViSi dataset from routine inpatient monitoring, we derived summary statistics for each patient and analyzed the summary statistics to find patterns in the patient population. This analysis revealed several types of patient posture habits. The careful analysis of summary statistics from many patients with different diseases forms a reference to further identify specific posture habits in patients with a certain disease. The procedure that we developed for data cleaning and preprocessing can have broad application to other monitoring systems in hospital use.