The wealth of routinely collected data in Neonatal Intensive Care Units (NICUs) has great potential. Morbidities, such as bronchopulmonary dysplasia, retinopathy of prematurity, and sepsis, can possibly be predicted when coupling analyses of routinely measured vital signs or derivatives with outcomes. One example is the HeRO symphony system predicting sepsis from variability of heart rate. (1) In real time, algorithms can summarize relevant data, detect anomalies, and notify bedside staff of risk factors in certain diseases. Routinely collected data could be used to develop algorithms or find associations, retrospectively.

However, it is unclear at what frequency data should be sampled. In our NICU, data are often sampled at least once per second (one-per-second data, i.e., 1 heart rate value per second) for prospective studies, but routinely collected vital parameters are only sampled once per minute (one-per-minute data). This keeps up performance of the clinical patient data management system, and prevents high costs associated with storage of data. Other NICUs may have similar infrastructure in place with data already collected and available. Although the data could be collected at a higher frequency, it is unclear whether lower frequency data are already enough.

We hypothesized that lower frequency data could, in some cases, be sufficient to run retrospective studies. In this short report, we investigated what to expect when using one-per-minute data abstracted from one-per-second data and investigated under what conditions one-per-minute data could be used.

Routinely collected data from a previous study were used; the ethical review committee of Leiden Den Haag Delft provided a statement of no objection for obtaining and publishing the anonymized data (G19.075). (2) Data recordings were included from infants born under 30 weeks of gestation in our tertiary-level perinatal center between 1 November 2018 and 15 March 2020. Recordings were excluded if they contained no data on peripherally measured oxygen saturation (SpO₂).

There were data available from 92 patients, with a median of 1,151,774 [577,843–2,586,608] one-per-second data points per patient. An excerpt from a data recording is shown in Figure 1. When processed to one-per-minute data, there were 19,462 [9,129–43,162] data points left. The time difference between the first and last entries in the data recording was 375 h, 24 min, and 46 s [157:59:33–762:23:11] for the one-per-second data, and 367 h, 58 min, and 30 s [155:11:45–757:12:00] for the one-per-minute data.

In the one-per-second data, the mean SpO₂ was 94.96 (1.88) vs. 94.96 (1.87) in the one-per-minute data (Table 1), and the standard deviation of SpO₂ was 3.14 (0.92) vs. 3.15 (0.91), respectively. SpO₂ was found to be within the target range in 70.96 [57.16–91.50]% of the time in the one-per-second data, and in 71.06 [57.00–91.53]% of the time in one-per-minute data. Hypoxic values under 80% were found in 0.36 [0.09–0.85]% of SpO₂ values in the one-per-second dataset vs. 0.35 [0.10–0.85]% of SpO₂ values for the one-per-minute dataset. Missing values were also similar [2.06 (1.59–2.91)% vs. 2.06 (1.52–2.87)%]. Bland-Altman plots can be found in the Supplementary Figures 1–6.

The per-patient-average-inspired FiO₂ was 22.65 [21.67–24.56] vs. 22.70 [21.69–24.59], and there was no difference in oxygen days [0 (0–2) in both datasets].

Sub analyses of groups with only less than 100 h (Table 2), 200 h, and half of the total dataset showed similar results with little difference between one-per-second and one-per-minute data.

In this study, we found little to no difference when comparing descriptive statistics of one-per-minute data and one-per-second data from the same source. This included clinically relevant outcomes as proportion of time within the oxygen saturation target range, hypoxia, and days of supplemental oxygen. Sub analyses of recording under 100 or 200 h showed no difference. The results suggest that routinely collected data recordings of comparable length or longer could be used for retrospective studies.

Although using routinely collected vital parameters for big data analysis and machine learning is increasingly popular, to our knowledge, there is no literature available describing the minimum data sampling frequency for our purpose. From the field of data signal processing, the Nyquist-Shannon sampling theorem (3) provides us a guideline for a sufficient sample rate, but this is aimed at reproducing the original signal, and not the summarizing statistic we often require for our retrospective studies.

One could argue that taking a sample every minute from continuous vital signs monitoring is somewhat analogous to research in general. It is uneconomical to study an entire population; thus, we take a representative sample. When the change of being sampled is related to the outcome, there is a chance of biased results. Although our sample is not at random, the value is always extracted in the first second of the minute. It is unlikely that a vital parameter like the heart rate is systematically lower in the first second of the minute or, in other words, related to the outcome. There may be a detectable circadian trend in the average heart rate, but the instantaneous heart rate should not be related to a certain second within a minute.

Limitations of our study are that one-per-minute data cannot be used to calculate the length of vital sign episodes, for example, the duration of a hypoxic episode, or other more elaborate outcomes from complex signal processing techniques. We have only investigated descriptive statistics of SpO₂ and FiO₂, and most of the data recordings had minimum duration of 100 h. It should also be noted that intra-recording differences were present, but these are averaged out over the entire set. Finally, to prevent synchronization issues, we did not compare our PDMS data directly with our higher frequency data, but down-sampled the latter. However, because in neither case, filtering, anti-aliasing or other processing is done, they are comparable.

In our study, descriptive statistics of lower frequency data were comparable to high frequency data and could be used for retrospective analyses. Comparable routinely collected one-per-minute data could be used to develop algorithms or find associations, retrospectively.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Medical Ethical Review Committee of Leiden Den Haag Delft. Written informed consent from the participants’ legal guardian/next of kin was not required to participate in this study in accordance with the national legislation and the institutional requirements.

HS was the principal author of the study and contributed to the design of the study, performed clinical data collection, verification and analyses, and drafted the first version of the manuscript. JD and RW drafted the first version of the manuscript. PD contributed to the design of the study and drafted the first version of the manuscript. SP performed data analyses and interpretation, and drafted the first version of the manuscript. AP contributed to the design of the study and data interpretation and drafted the first version of the manuscript. All authors provided substantial intellectual contributions and approved the final version of the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.