Eligibility criteria and search

Data collection and analysis

Search

The result of the search is presented in the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) flow diagram (Fig 1). A total of 25,921 records were identified from the two initial searches. After removing duplicates, 10,320 records were screened and 9,999 were considered irrelevant. From the remaining 321 records, 20 were not retrievable as full-text articles despite intensive use of interlibrary loan. Three-hundred-one records were assessed for eligibility by full-text screening leading to the exclusion of 250 records with reasons. Three studies (non-RCTs and <1.000 patients) were excluded due to language restriction. Three studies were classified as ‘ongoing’ (none of them completed) and 15 studies remained ‘awaiting classification’ due to insufficient information on eligibility criteria. A search update for RCTs on 26 October 2024 identified further 218 records, from which one study was deemed eligible after screening. A second update search for RCTs conducted on 29 September 2025 identified 408 additional records; however, none were eligible for inclusion. Finally, 26 studies were included in this systematic review [17–49]. References of studies ongoing, excluded, and ‘awaiting classification’ along with reasons are presented in Supplement (S1 Table, S2 Table, S3 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram.

Results of the searches are presented in the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) flow diagram. Abbreviations: Randomized controlled trial (RCT).

https://doi.org/10.1371/journal.pdig.0001126.g001

Study characteristics and settings

One study utilized a traditional cRCT design, in which ICUs were randomized to either the telemedicine intervention or SoC group [41]. Two cRCTs used a stepped-wedge design, in which all ICU clusters transitioned from a control phase to the telemedicine phase [31,45]. Randomization determined the timing of each ICU’s transition. The majority of the 23 non-randomized studies used a before-and-after study design (n = 18). Additionally, four cohort studies were included [20,38,43,47], and one case-control study [27] (Table 1). Twenty-one studies were conducted in the United States, two in Germany [31,45], one in Australia [39], one in Brazil [41], and one did not report the country of study conduct [23] (Table 1). Seven out of 26 studies did not report a funding source at all, while three studies were fully or partially funded by industry, and the remaining studies by government, department, or a foundation (Table 1). All included studies compared telemedicine in intensive care to SoC without telemedicine (Table 1); however, in most studies SoC description was insufficient or completely absent. In 11 studies tele-centers provided telemedicine exclusively while in seven studies one or more hospitals acted as telemedicine provider. In three studies a combination of both was reported and in five studies no telemedicine provider information was available (Table 1). Recipient hospitals in eligible studies were community/rural (n = 5), urban/tertiary (n = 3), or university hospitals (n = 2). Mostly, a combination of different hospital types was reported (n = 13) and in three studies no recipient information was available (Table 1). Many hospitals used high-tech audio-video communication tools and EHR with automated data transfer (n = 15), while only three studies reported the use of low-tech equipment along with other approaches (n = 8) (S4 Table). The mode of communication varied widely across included studies, ranging from unstructured interviews during contact on demand only (n = 1) to combinations of contact on demand, emergency contact, together with daily rounding (n = 5), as well as structured interviews during contact on demand and daily rounding (n = 2), among other sorts of variations (n = 18) (S4 Table). Reporting on the expertise in telemedicine of the provider or recipient, including additional training or implementation aids, was generally inadequate (S4 Table). In most included studies, telemedical consultation combined with therapeutic decision making was the primary mode of delegation for remote caregivers (n = 12), followed by consultation only (n = 7), and therapeutic decision making only (n = 1) (S4 Table). Four studies did not report this information and two had additional regulations. Further information on telemedical technical equipment, mode of communication, and delegations grade are provided in Supplement (S4 Table). Twenty-four of the 26 included studies comprised 2,164,508 analysed patients, with a median (range) number of 7,314 (525-1,123,563) (Table 1). One NRSI reported the amount of telemedical consultations only [21]. Twenty of the included studies reported a mean or median age of > 60 years for the patient population, while two studies did not report the populations’ age at all (S5 Table). Most of the studies included a higher proportion of men than women; however, the difference was not significant within individual studies (S6 Table). Three of the included studies utilized an overlapping pool of study participants in their analyses: one with 6,988 enrolled patients [34], another with 563,491 enrolled patients [24], and a third without reporting the number of enrolled patients [38] (Table 1). Notably, more than 90% of the participants in this specific pool were men [24,34,38] (S6 Table). To avoid duplication, only one study [34] was used for further analyses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Study characteristics of included studies.

https://doi.org/10.1371/journal.pdig.0001126.t001

Three RCTs, one with cluster design and two with stepped-wedge cluster design, reported data of our main outcome set used in this systematic review [31,41,45]. Spies 2023 [45] reported ICU mortality, overall-mortality at 180 days, quality of life at 180 days, and ICU LOS. Marx 2022 [31] reported ICU mortality, and Pereira 2024 [41] ICU mortality and ICU LOS. Additionally, eight NRSIs reported outcomes of our main outcome set. ICU mortality was analysed by seven studies [28,29,34,39,44,46]. Three studies reported ICU LOS [28,34,49], and two studies overall mortality at 30 and 90 days [27,34]. Adjustments for different confounders are shown in Tables 2 and 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summary of findings: Telemedicine vs SoC in ICU patients ((sw-)cRCTs).

https://doi.org/10.1371/journal.pdig.0001126.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of findings: Telemedicine vs SoC in ICU patients (NRSIs).

https://doi.org/10.1371/journal.pdig.0001126.t003

We contacted 24 authors from 22 different studies regarding study characteristics and missing or unclear data. Four authors provided additional information about unclear data and patient details [24,39,45,48].

Risk of bias of included studies

The risk of bias of the main outcomes ICU mortality, overall mortality at longest follow-up, ICU LOS, and quality of life were separately assessed in three different cRCTs [31,41,45] by using the RoB 2 tool [10] (S7 Table). Spies 2023 [45] was rated as overall high risk of bias for all outcomes due to recruitment bias leading to selection of sicker patients to the telemedicine group. Additionally, all outcomes, except quality of life, were not adjusted for time trends in this study. ICU mortality in Marx 2022 [31] was assessed as overall high risk of bias due to deviations from intended intervention as sicker patients in the telemedicine group preferably received the intervention, combined with bias for missing adjustment of time trends. ICU LOS overall risk of bias in Pereira 2024 [41] was rated as ‘some concerns’ due to bias in measurement of the outcome, and ICU mortality was rated as ‘no concerns’ (S7 Table).

We assessed the main outcomes ICU mortality, overall mortality at longest follow-up, and ICU LOS in eight different NRSIs [27–29,34,39,44,46,49] using ROBINS-I [12] (S8 Table). ICU mortality was assessed as overall moderate risk of bias in one study [28], serious in three studies [28,44,46], and critical in three studies [29,34,39]. Both studies, contributing overall mortality at longest follow-up, received an overall critical risk of bias [27,34]. The overall risk of bias for ICU LOS was rated as moderate [28], serious [49], and critical [34] in one study each (S8 Table). Problematic studies most frequently revealed bias in selection of participants and bias due to missing data.

Effects of interventions: primary outcomes

Mortality on intensive care units.

Two sw-cRCTs involving 5,915 participants reported ICU mortality but were not pooled in a meta-analysis due to clinical and methodological heterogeneity [31,45]. Additionally, one cRCT involving 15,230 participants reported ICU mortality [41]. Spies 2023 (RR 0.89 (95% CI 0.54–1.46)) and Pereira 2024 (RR 1.09 (95% CI 0.94–1.26)) showed no evidence for a difference between telemedicine and SoC, while evidence in Marx 2022 (RR 2.29 (95% CI 1.51 – 3.48)) favoured SoC. In both sw-cRCTs, selection of considerably sicker patients for telemedicine likely biased the observed results. This assumption is supported by disease severity adjusted data for ICU mortality (OR 1.28, 95% CI 0.91–1.79)) showing no evidence for a difference between groups [31]. Certainty of evidence for Marx 2022 and Spies 2023 was very low due to very serious risk of bias and serious inconsistency, while for Pereira 2024 certainty of evidence was very low due to extremely serious imprecision (Table 2, Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Effect direction plot for RCTs.

Effect direction plot summarizing findings by showing the direction of effects across RCTs. Abbreviations: Cluster randomized controlled trial (cRCT), intensive care unit (ICU), length of stay (LOS), randomized controlled trial (RCT), standard of care (SoC), stepped-wedge cluster randomized controlled trial (sw-cRCT). Legend: ▲ telemedicine increases outcome. ◄► telemedicine has no to minimal effect on outcome.

https://doi.org/10.1371/journal.pdig.0001126.g002

Six NRSIs reported adjusted effect estimates for ICU mortality involving 136,125 participants. Studies were not combined in meta-analysis due to clinical and methodological heterogeneity [28,29,34,39,44,46]. Evidence from three studies favoured telemedicine [28,29,44], while the others showed no evidence for a difference between telemedicine and SoC [34,39,46]. Certainty of evidence was very low due to very serious risk of bias and serious inconsistency (Table 3, Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Effect direction plot for NRSIs.

Effect direction plot summarizing findings by showing the direction of effects across NRSIs. Abbreviations: Intensive care unit (ICU), length of stay (LOS), non-randomized study of intervention (NRSI), standard of care (SoC). Legend: ▲ telemedicine increases outcome. ▼ telemedicine decreases outcome. ◄► telemedicine has no to minimal effect on outcome.

https://doi.org/10.1371/journal.pdig.0001126.g003

Overall mortality at longest follow-up

One sw-cRCT reported overall mortality at 180 days (n = 1,462), which showed no evidence for a difference between telemedicine and SoC (RR 1.02 (95% CI 0.76–1.36)) [45]. Certainty of evidence was very low due to very serious risk of bias and serious imprecision (Table 2, Fig 2).

Two NRSIs (n = 295,991) reported overall mortality at longest follow-up but were not pooled due to clinical and methodological heterogeneity. One study showed no evidence for a difference between telemedicine and SoC for 30-day mortality [34] and another favoured SoC for 90-day mortality [27]. Certainty of evidence was very low due to very serious risk of bias and serious imprecision (Table 3, Fig 3).

Length of stay on intensive care units

Length of stay on ICUs was reported by one cRCT (mean (SD) 8.1 (10) days (telemedicine) vs 7.1 (9) days (SoC)) and one sw-cRCT (median [IQR] 6 [4–13] days (telemedicine) vs 5 [3–11] days (SoC)) (n = 16,692) [41,45]. Both studies showed no evidence for a difference between telemedicine and SoC. For Spies 2023, certainty of evidence was very low due to very serious risk of bias and serious imprecision thus the evidence is very uncertain about the effect of telemedicine on ICU LOS. Certainty of evidence for Pereira 2024 was moderate due to serious risk of bias consequently telemedicine probably has little or no difference on ICU LOS (Table 2, Fig 2).

Three NRSIs (n = 21,933) reported adjusted data on ICU LOS but were not pooled due to clinical and methodological heterogeneity [28,34,49]. The studies provided conflicting results: data from two studies favoured telemedicine [28,49], while one study showed no evidence for a difference between telemedicine and SoC [34]. Certainty of evidence was very low due to very serious risk of bias and serious inconsistency (Table 3, Fig 3).

Quality of life (at 180 day)

Quality of life, measured in one stepped-wedge cluster-RCT (n = 786) using the EQ-5D-5L, showed no evidence for a difference between telemedicne and SoC (MD -2.71 (95% CI 6.95 fewer to 1.53 more)) [45]. Certainty of evidence was very low due to very serious risk of bias and serious imprecision (Table 2, Fig 2).

Effects of interventions: secondary outcomes

None of the included studies addressed disease-related detection rate, acceptance, changes in therapeutic goals, or triage outcomes. Studies investigating hospital mortality and hospital LOS reported heterogeneous results (S9 Table, S10 Table). Adherence to best practice guidelines was notably higher in telemedicine groups [31] (S11 Table). The same study demonstrated increased compliance with sepsis bundles and timely antibiotic administration in the telemedicine group [31] (S11 Table). Additionally, this sw-cRCT found a slight reduction in antibiotic days with telemedicine [31] (S12 Table). One RCT noted fewer ventilator-free days at 28 days in telemedicine settings [41], while another reported a longer duration of mechanical ventilation for telemedicine patients [45] (S12 Table). Transfer rate reported in two studies [31,38] was higher in telemedicine groups (S13 Table). One sw-cRCT reported improved fulfilment of a subset of process and quality indicators under telemedicine management [45] (S14 Table). However, interpretation of the clinical relevance of this improvement is challenging due to high baseline adherence in some ICU clusters.

Subgroup analyses

None of the studies compared different types or modes of telemedicine in intensive care settings.

One study reported ICU mortality for sepsis patients with uncertain effect (OR 0.68 (95% CI: 0.23–10.87) [31].

Methodological limitations and heterogeneity

Several sources of bias and heterogeneity complicate the interpretation of the evidence. Two sw-cRCTs were prone to selection bias, where telemedicine was preferentially applied to sicker ICU patients, and likely distorting patient-relevant outcomes such as mortality, LOS and quality of life. In several NRSIs, it also remained unclear whether all patients in the intervention group consistently received telemedicine, further increasing the risk of bias [29,39,44,46]. Additionally, hospitals involved in these studies varied widely in ICU or hospital settings, staff allocation, and the implementation of telemedicine interventions leading to high clinical heterogeneity and limited comparability.

Moreover, the differences in authorizations of the telemedicine teams in terms of consultation vs decision making and treatment responsibility presumably leads to a high variability in the implementation of telemedical recommendations in clinical practice. Only one included study reported full authority to the telemedicine provider [23], which may have a more significant impact on patient-relevant outcomes than consultation-only models.

The same applies for the availability of telemedicine providers and their telemedicine expertise as well as the access to training programs, which are also crucial factors. High heterogeneity among studies was also found on analytical level. Adjustments for confounders were inconsistent or entirely absent in many NRSIs [19,20,24,25,27,29,32,33,39,40,43,46–48].

Another contributing factor to uncertainty is that many included studies were designed and powered for different primary outcomes than those analyzed in our systematic review [21,23,24,30,31,40,45,48], resulting in suboptimal study designs or analyses for our research questions. As previously noted in another systematic review [50], it often remains unclear whether telemedicine interventions aim to maintain existing standards of care or to actively improve outcomes — further complicating the interpretation of study objectives and findings.

This pronounced heterogeneity across studies underscores the multifaceted nature of telemedicine as a complex intervention, reflecting its variability in design, delivery, and integration into clinical workflows as well as the degree of personal motivation or engagement among providers and recipients. These inconsistencies collectively challenge the ability to pool results effectively and to draw definitive conclusions about telemedicine ‘s impact on patient-relevant outcomes in intensive care.

Agreements and disagreements with other studies or systematic reviews

These outcome variations are not unique to our systematic review. Other reviews have likewise observed inconsistent effects in primary studies [4–7], largely attributable to different primary study designs (mainly NRSIs) and differing analytical approaches compared with a strong heterogeneity regarding hospital and ICU characteristics, e.g., the technical implementation, staff allocation, and the delegations’ grade. Kalvelage 2021 [7] pooled adjusted and unadjusted outcome data from NRSIs in meta-analyses, while Chen 2018 [4] and Wilcox 2012 [5] relied exclusively on unadjusted outcome data. Mackintosh 2016 included only two NRSIs, both assessed as high risk of bias [51]. Collectively, a plenty of systematic reviews call attention to the high heterogeneity and reporting problems among studies regarding clinical settings and methodology. However, none of the published systematic reviews opted against pooling due to heterogeneous primary studies, as recommended by Cochrane [52]. Meta-analyses of very diverse studies can be misleading and should only be considered when a group of studies is sufficiently homogeneous in terms of participants, interventions, and outcomes to provide a meaningful summary [52]. A strength of our systematic review is that we chose not to pool the studies, given their heterogeneity. This lack of unity among literature highlights the importance of evidence-based strategies for telemedicine implementation in intensive care. To our knowledge, this is the first systematic review including evidence from (sw-)cRCTs providing a higher level of evidence compared to NRSIs.

Patient and public involvement

Patients or the public were not involved in the design, conduct, reporting, or dissemination plans of our research.