How Robust are the Evidences that Formulate Surviving Sepsis Guidelines? An Analysis of Fragility and Reverse Fragility of Randomized Controlled Trials that were Referred in these Guidelines.

OBJECTIVES: "Surviving Sepsis Campaign: International Guidelines for Management of Sepsis and Septic Shock: 2016" provides guidelines in regard to prompt management and resuscitation of sepsis or septic shock. The study is aimed to assess the robustness of randomized controlled trials (RCTs) that formulate these guidelines in terms of fragility index and reverse fragility index.
METHOD: RCTs that contributed to these guidelines having parallel two-group design, 1:1 allocation ratio, and at least one dichotomous outcome were included in the study. The median fragility index was calculated for RCTs with significant statistical outcomes, whereas the median reverse fragility index was calculated for RCTs with nonsignificant statistical results.
RESULTS: Hundred RCTs that met the inclusion criteria were analyzed. The median fragility index was 5.5 [95% confidence interval (CI) 1-30] and median reverse fragility index was 13 (95% CI 12.07-16.8) at a p value of 0.05. The median reverse fragility index was 16 (95% CI 10-26) at a p value of 0.01. Most of the RCTs included in this analysis were of good quality, having a median Jadad score of 6.
CONCLUSION: This analysis found that the surviving sepsis guidelines were based on highly robust RCTs with statistically insignificant results and on some moderately robust RCTs with statistically significant results. RCTs with statistically insignificant results were more robust than RCTs with statistically significant results in regard to these guidelines. HIGHLIGHTS: The study assessed the robustness of randomized controlled trials (RCTs) that were used to formulate surviving sepsis guidelines. Most RCTs showed statistically nonsignificant results. RCTs with statistically significant results were moderately fragile whereas RCTs with nonsignificant results were more robust. HOW TO CITE THIS ARTICLE: Choupoo NS, Das SK, Saikia P, Dey S, Ray S. How Robust are the Evidences that Formulate Surviving Sepsis Guidelines? An Analysis of Fragility and Reverse Fragility of Randomized Controlled Trials that were Referred in these Guidelines. Indian J Crit Care Med 2021;25(7):773-779.

Introduction

The probability values, more popularly known as p values, are widely used to quantify the statistical significance of observed results. The practice of significance testing originated from the concept and practice of the renowned statistician, R.A. Fisher, in the third decade of the 20th century.[1] However, p values have been frequently subjected to criticism due to its potential misinterpretation. When a p value was introduced, it was not supposed to be used as a definitive test but was a casual way to determine whether the evidence was significant in an old-fashioned way. It is often assumed that a lower p value indicates a more statistically significant result. Many erroneously regard statistical significance as having clinical significance. This is oversimplification and may result in overemphasis on the clinical importance of the study. A large study could have the same p value as a very small study. While both are regarded as “statistically significant,” the p value does not provide any indication that there is a clear distinction between these studies, leading one to conclude that the likelihood of a true effect is the same. Another important fallacy is that only one event can make a significant result nonsignificant and vice versa. The former is typically interpreted as indicating a more important treatment effect, although there being minimum absolute difference between the two types of result.[2,3]

Therefore, to decrease the absolute reliance on p value, various measures have been postulated, and they are lowering p value threshold, using alternative approaches like effect size and confidence interval, Bayes factor, Akaike information criterion, incorporation of fragility index (FI), etc.[4-6] The concept of fragility was introduced by Feinstein in the epidemiology literature.[7]

This implies the minimum number of patients whose status would have to be changed from a “nonevent” to an “event” in order to turn a statistically significant result into a nonsignificant result.[7] If lesser numbers are required to change the statistical significance of the study, it is regarded to be the lack of robustness of a trial result. FI is exclusively applied to trials that reach traditional statistical significance. To check the robustness of a statistically nonsignificant trial, reverse fragility index (RFI) has been used.[8] RFI provides a measure of robustness in the neutrality of results when assessed from a clinical perspective.

“Surviving Sepsis Campaign: International Guidelines for management of Sepsis and Septic Shock: 2016” provided 93 statements on early management and resuscitation of patients with sepsis or septic shock.[9] These guidelines are a careful synthesis of available randomized controlled trials (RCTs), systematic review and meta-analysis, and case-control studies that encompass a wide range of management strategies including early resuscitation, goal-directed therapy, antibiotic therapy, fluid therapy, vasoactive medications, corticosteroids, immunoglobulins, blood purifications, anticoagulants, mechanical ventilation, sedation analgesia, glucose control, renal replacement therapy, etc.[9] The purpose of this study is to apply FI and RFI analysis to the latest surviving sepsis guidelines (SSG) and to assess the fragility of RCTs, reporting dichotomous outcome parameters.

Materials and Methods

Data Search

Recent Surviving Sepsis Campaign guidelines published in the year 2016 were reviewed. Two independent investigators (SKD and NSG) screened all the RCTs referenced in guidelines and assessed them for inclusion. Any disagreement was resolved by consensus with a third author (PS).

RCTs with parallel two-group design

1:1 allocation ratio

At least one dichotomous outcome was included in the study.

Eligibility Criteria

Letters, editorials, systematic reviews or meta-analyses, opinions, observational studies, economic or cost-effective analyses of RCTs, cohort nonrandomized studies, and quasi-randomized trials were excluded.

Data Collection

A prespecified data collection form was used to extract the following data from all RCTs: studied intervention, authors, binary outcomes, sample sizes, number of patients with events, and number of patients without events. We prioritized the primary outcomes for the analysis; however, when analyzable data were not available, secondary dichotomous outcomes related to mortality were included.

Quality Assessment

Quality assessment of included studies was done by one investigator (PS) using “modified Jadad scale.” A questionnaire based eight questions was used to assess randomization, blinding withdrawal or dropouts, description of inclusion/exclusion criteria, assessment of adverse effects, and description of the statistical plan. A score of 1 to 8 was given to each study where 8 denotes maximum robustness whereas 1 denotes least.[10]

Outcomes

The outcomes were FI and RFI at p values of 0.05 and 0.01, fragility quotient (FQ) and reverse fragility quotient (RFQ).

Statistical Analysis

For each included outcome from RCTs, a two-by-two contingency table was created. FI was calculated according to the method described by Walsh et al.[11] The number of events was added to a group with a smaller number of events while subtracting nonevents from the same group to keep the total number of participants constant. Events were added iteratively and calculations were done with a Fisher's exact test for each addition until the calculated p value became just more than 0.05. RFI was calculated according to the method described in a recent publication.[8] The RFI was calculated by subtracting events from the group with a lower number of events while simultaneously adding nonevents to the same group to keep the number of participants constant until the Fisher's exact test two-sided p value became less than 0.05.[8] A similar method was used to calculate RFI at a p value of 0.01.

FI or RFI is an absolute measure of stability, irrespective of trial size. We analyzed FQ and RFQ as a relative measure of fragility. This was calculated by dividing the FI or RFI by its respective sample size.[12]

Subgroup analysis was done to analyses FI and RFI of studies testing similar domains of sepsis management, e.g. studies dealt with mechanical ventilation.

FI was calculated using the online FI calculator www.clincalc.com. To calculate a Fisher's exact test two-sided p value, the online calculator https://www.graphpad.com/quickcalcs was used.

Result

After screening 655 references of surviving sepsis guidelines 2016 (SSG2016), a total of 201 RCTs were identified. Of these, 100 RCTs were included in the final analysis. Among the included RCTs, 22 had dichotomous statistically significant outcome measures and 78 studies reported statistically insignificant dichotomous outcome measures (Fig. 1). Median sample size of RCTs with significant result was 286 [95% confidence interval (CI) 32–6,104]. The median sample size of RCTs with statistically insignificant results was 520 (95% CI 31–6,997) (Tables 1 and 2).

Fig. 1

Review process and included studies

Table 1

Characteristics of included studies with statistically significant results

Studies	Intervention	Sample size	Fragility index	Fragility quotient	Jadad score
Rivers E	EGDT	263	4	0.01	7.5
Bernard GR	Recombinant human protein C	1,690	15	0.008	8
de Jong E	Procalcitonin-guided antibiotic therapy	1,546	9	0.005	6
Martin C	Dopamine vs norepinephrine	32	5	0.15	5
Corwin HL	Recombinant erythropoietin	1,302	30	0.20	8
Bollaert PE	Hydrocortisone	41	7	0.17	—
Amato MB	Protective ventilation	53	1	0.01	6
Brower RG	Low tidal volume	861	12	0.01	5
Villar J	High PEEP, low tidal volume	103	1	0.009	5
Guérin C	Prone position 14	466	20	0.04	6
Peek GJ	ECMO	180	2	0.01	6
Ferguson ND	HFOV	548	10	0.01	6
Ferrer M	NIV	105	4	0.03	5
Gao Smith F	Intravenous β2 agonist in ARDS	326	2	0.006	8
Futier E	Intraoperative low tidal volume	400	17	0.04	8
Drakulovic MB	Supine body position	86	3	0.03	5
Schweickert WD	Early physical and occupational therapy	104	3	0.02	6
van den Berghe G	Intensive insulin therapy	1,548	7	0.004	6
Finfer S	Intensive insulin therapy	6,104	9	0.001	6
Fuentes-Orozco C	L-alanyl-L-glutamine	33	3	0.09	8
Detering KM	Advance care planning on end-of-life care	309	6	0.01	5
Aguado JM	Galactomannan and PCR-based DNA detection of aspergillus	203	1	0.004	6

EGDT, early goal-directed therapy; ECMO, extracorporeal membrane oxygenator; HFOV, high-frequency oscillating ventilation; NIV, noninvasive ventilation

Table 2

Characteristics of included studies with nonsignificant statistical results

Author	Intervention	Sample size	Reverse FI at p <0.5	Reverse FI at p <0.01	Fragility quotient	Jadad score
Peake SL	Goal-directed resuscitation	1,591	28	35	0.01	6
Yealy DM	EGDT	895	14	20	0.01	6
Mouncey PR	EGDT	1,260	29	36	0.02	6
Hayes MA	Elevation of oxygen delivery by dobutamine	100	1	3	0.005	6
Jansen TC	Lactate-guided resuscitation	348	2	7	0.005	6
Jones AE	Lactate vs ScvO2-guided resuscitation	300	6	8	0.02	6
Lyu X	Lactate clearance	100	6	8	0.06	—
Brunkhorst FM	Moxifloxacin and meropenem vs meropenem	600	*13,12	18,19	0.02,0.02	6
Chastre J	Eight vs 15 days of antibiotic therapy	401	12	15	0.03	8
Sawyer RG	Short-course antimicrobial therapy	517	17	23	0.03	6
Dunbar LM	Levofloxacin 750 mg vs 500 mg	528	18	25	0.03	8
Hepburn MJ	Short-course antimicrobial therapy	87	7	14	0.08	8
Rattan R	Antibiotic duration	112	7	8	0.06	6
Caironi P	Albumin vs crystalloid	1,818	36	45	0.02	6
Russell JA	Vasopressin norepinephrine	781	12	18	0.01	8
Gordon AC	Vasopressin norepinephrine	408	19	24	0.04	8
De Backer D	Dopamine vs norepinephrine	1,679	21	35	0.004	8
Annane D	Epinephrine vs norepinephrine plus dobutamine	330	12	16	0.03	8
Gordon AC	Levosimendan	516	10	14	0.02	8
Briegel J	Hydrocortisone	40	5	8	0.1	—
Sprung CL	Hydrocortisone	233	11	13	0.04	8
Annane D	Hydrocortisone and fludrocortisone	299	10	12	0.03	8
Huh JW	Corticosteroids	130	11	12	0.07	6
Keh D	Corticosteroids	340	13	15	0.03	8
Holst LB	Transfusion threshold	998	22	30	0.02	7.5
Zumberg MS	Platelet transfusion	159	6	8	0.04	5
Stanworth SJ	Platelet transfusion	600	2	8	0.02	6
Werdan K	Immunoglobulin G	624	18	23	0.03	7
Payen DM	Polymyxin hemoperfusion	243	10	12	0.04	6
Livigni S	Plasma filtration adsorption	184	12	15	0.07	6
Warren BL	Antithrombin III	2,314	46	58	0.02	8
Vincent JL	Thrombomodulin	741	7	12	0.02	8
Ranieri VM	Drotrecogin alfa	1,680	17	25	0.01	8
Papazian L	Cisatracurium infusion in ARDS	339	4	6	0.02	8
Brochard L	Reduction of tidal volume	116	7	9	0.06	6
Brower RG	Lower PEEP vs higher PEEP	549	13	18	0.02	5
Mercat A	PEEP	767	13	17	0.02	6
Guerin C	Prone position	791	22	28	0.03	6
Young D	HFOV	795	25	30	0.03	6
Meade MO	Low TV, recruitment maneuvers, and high PEEP	983	11	18	0.01	6
Antonelli M	NIV	64	6		0.09	5
Frat JP	HFNC	200	6	9	0.03	6
Wiedemann HP	Conservative vs liberal fluid management	1,000	14	20	0.01	6
Wheeler AP	PAC vs CVC	1,001	21	27	0.02	—
Richard C	Pulmonary artery catheter	676	21	26	0.02	6
Harvey S	Pulmonary artery catheter	1,041	17	22	0.02	6
Rhodes A	Pulmonary artery catheter	201	14	18	0.07	6
Sandham JD	Pulmonary artery catheter	1,996	22	28	0.01	6
van Nieuwenhoven CA	Semirecumbent position	221	4	5	0.01	6
Van den Berghe G	Intensive insulin therapy	1,200	17	25	0.01	6
Arabi YM	Intensive insulin therapy	523	8	10	0.01	6
Brunkhorst FM	Insulin therapy and pentastarch resuscitation	537	15	20	0.02	4
De La Rosa Gdel C	Strict glycemic control	504	11	16	0.02	6
Kalfon P	Intensive insulin therapy	2,666	25	35	0.01	6
Preiser JC	Intensive insulin therapy	1,101	15	19	0.01	6
Augustine JJ	Continuous vs intermittent dialysis	80	11	16	0.13	5
Mehta RL	CRRT vs IHD	164	13	15	0.07	6
Uehlinger DE	CRRT vs IHD	125	10	15	0.08	6
Vinsonneau C	CRRT vs IHD	359	16	22	0.05	6
Bellomo R	Intensity of CRRT	1,464	39	44	0.02	5
Palevsky PM	Intensity of CRRT	1,124	22	30	0.02	6
Gaudry S	Timing of RRT	619	21	26	0.04	6
Zarbock A	Timing of RRT	231	5	9	0.02	6
Cook D	Dalteparin vs unfractionated heparin	3,746	15	21	0.004	6
Harvey SE	Enteral vs parenteral nutrition	2,388	31	40	0.01	6
Doig GS	Early parenteral nutrition	1,372	22	27	0.01	7.5
Arabi YM	Permissive underfeeding	894	20	25	0.02	6
Singh G	Postoperative enteral feeding	43	7	8	0.16	4
Petros S	Hypo vs normocaloric	100	1	2	0.02	6
Reignier J	Not monitoring gastric residual volume	449	13	16	0.02	6
Valenta J	High-dose selenium	150	7	9	0.04	4
Caparrós T	High-protein diet enriched with arginine, fiber, antioxidant	220	4	7	0.03	7.5
Kieft H	Immunonutrition	597	17	26	0.03	8
Grau T	Immunonutrition	127	8	10	0.07	8
Galbán C	Immune-enhancing diet	176	1		0.03	6
Puskarich MA	L carnitine	31	5	6	0.19	8
Young P	Buffered crystalloid vs saline	2,092	21	28	0.01	8
Finfer S	Albumin vs saline	6,997	65	80	0.09	8

EGDT, early goal-directed therapy; HFOV, high-frequency oscillating ventilation; NIV, noninvasive ventilation; PEEP, positive end-expiratory pressure; PAC, pulmonary artery catheter; CRRT, continuous renal replacement therapy; IHD, intermittent hemodialysis

Median FI was 5.5 (95% CI 1–30) and median RFI was 13 (95% CI 12.07–16.8) at a p value of 0.05.

Median FQ was 0.01 (95% CI 0.01–0.02) and median RFQ was 0.02 (95% CI 0.02–0.04)

Median RFI was 16 (95% CI 10–26) at a p value of 0.01.

Most of the RCTs included in this analysis were of good quality. The median Jadad score of RCTs with significant results was 6 (95% CI 5–8) and the median Jadad score of RCTs with nonsignificant results was also 6 (95% CI 4–8).

Subgroup Analysis

RCTs that are included in this analysis were grouped according to the domains they dealt with (Table 3). Three most commonly studied subjects that were analyzed by the RCTs were mechanical ventilation, nutrition, and goal-directed therapy. Fifteen studies were done on various ventilator strategies; ECMO and other supportive measures had a median FI and RFI of 4 and 12, respectively. Thirteen studies on nutrition were analyzed; of which 12 studies showed nonsignificant results having a median RFI of 7.5. Eight studies were done on the efficacy of goal-directed therapy; except one all RCTs had nonsignificant results with a median RFI of 6. Subgroup analysis also revealed that studies with insignificant results were more robust than those with significant results.

Table 3

Subgroup analysis of RCTs according to domains they dealt with

Subject	Studies with significant results	Studies with nonsignificant results	FI	FQ	RFI	RFQ
EGDT/GDT	1	7	4	0.01	6	0.02
Vasopressors/inotropes	1	5	5	0.15	10	0.02
Infection	2	6	5	0.0045	12	0.03
Ventilation, ECMO, and others related to oxygenation	7	8	4	0.01	12	0.03
Nutrition	1	12	3	0.09	7.5	0.03
Steroids	1	5	7	0.17	11	0.05
Adjunct therapy	1	6	15	0.008	17.5	0.025
Insulin therapy	2	6	8	0.002	15	0.01
Transfusion	—	3	—	—	6	0.02
Anticoagulant/DVT prophylaxis	—	1	—	—	15	0.004
Renal replacement therapy	—	8	—	—	16	0.03
Patient position	1	1	3	0.02	4	0.03
Pulmonary artery catheter	—	5	—	—	21	0.03
Intravenous fluids	—	3	—	—	40	0.03
End-of-life care	1	0	6	0.01	—	—
Physical therapy	1	0	3	0.02	—	—
Others	1	1	30	0.2	8	0.02

Discussion

This retrospective analysis of evidences that formulated SSG found that the guidelines are based on highly robust RCTs with statistically insignificant results and on some moderately robust RCTs with statistically significant results. The median sample size was larger in RCTs having nonsignificant statistical results.

FI has been evaluated on studies of anticancer medicines, heart failure, anesthesiology, and several other areas of biomedical science in order to assess the robustness of findings amid concern over the reproducibility of research.[13-23] A retrospective analysis calculated a median FI of 56 RCTs in critical care medicine reporting mortality. The median FI was 2 with an interquartile range (IQR) of 1 to 35.[24] Similar to our study, several clinical guidelines were subjected to FI analysis. An analysis of 32 RCTs included in the American College of Gastroenterology Guidelines of Crohn's disease reported a median FI of 3.[25] An analysis of 21 RCTs that were used to support treatment recommendations in the 2016 “Chest Guideline and Expert Panel Report on Antithrombotic Therapy for VTE Disease” found a median FI score of 5 (1–9).[26] Another study of 35 RCTs in the 2017 diabetes treatment guidelines reported that the median FI score was 16 (4–29).[27] Analysis of 25 RCTs in heart failure reported a median FI score of 26 (0–118).[16] Compared to these guidelines, RCTs of SSG had moderate robustness having a median FI of 5.5. Although there is no established cutoff value for FI or RFI as being robust or fragile, it is reasonable to postulate that the higher the value, the more “confidence” is on the possibility of the observed result to be robust. Studies that evaluated RCTs of various specialties reported median FI in the range of 2 to 26.[13-15,17,24] A study calculated FI of 399 RCTs published in NEJM, JAMA, The Lancet, BMJ, and Annals of Internal Medicine. Median FI was 8 with an IQR of 0 to 109.[11] The concept of RFI is relatively new. A recent study that analyzed 167 RCTs with statistically insignificant results that were published in NEJM, The Lancet, and JAMA reported a median RFI of 8 (5–13) at a p value of 0.05, which was lower than the median RFI of survival sepsis guidelines 2016.[8]

The FI and RFI are powerful and intuitive statistical concepts. They provide a useful additional tool for clinicians to use in assessing the treatment effect on patient outcomes. FI or RFI can help researchers to identify trials that are at risk of being overturned by future studies and avoiding overestimation of the significance of RCT results. However, looking at FI or RFI, it has been kept in consideration that many factors may influence them; of which, sample size, event rates, significant level, and statistical methods of association are important.[28]

The initial SSC guidelines were first published in 2004.[29] Since then, it has changed clinical behavior, improved quality of care, and decreased mortality in patients with severe sepsis and septic shock. The studies demonstrated that increased compliance was associated with a 25% relative risk reduction in mortality rate.[30] To our knowledge, analysis of FI and RFI of RCTs of these landmark guidelines was not done before. The present study may be first of its kind to assess the robustness of evidences that have shaped the guidelines. Previous studies appraising various clinical guidelines focused only on RCTs with significant results. Our study for the first time analyzed guidelines in regard to its RCTs with statistically insignificant results and also demonstrated that in these guidelines, RCTs with insignificant results are more robust than RCTs with statistically significant results.

Like any other statistical parameters, FI and RFI have also their own limitations. It can be used only to RCTs with dichotomous outcomes and 1:1 parallel study. RCTs with continuous outcomes cannot be evaluated. They do not account for the time at which events occurred which is a very important consideration, especially in oncological research.[31] FI alone does not convey a measure of precision so it has to be read in conjunction with the p value, sample size, CI, and number lost to follow-up. Because of these limitations, the present study could not analyze less than half of the RCTs included in SSG.

This is to be noted that clinical decision about the effectiveness of harm of an intervention should not be merely based on the statistical significance or lack of it.[32] Rather, it should be based on the magnitude of the treatment effect.[32] The statistical significance merely tries to quantify the probability of observing the reported effect size. FI and RFI do not quantify the treatment effect; rather, they can be used to understand the fragility of the probability of the treatment effect reported.

This analysis of 100 RCTs that contributed to SSG found a median FI of 5.5 and a median RFI of 13. Most RCTs had statistically nonsignificant results, and they are more robust than statistically significant studies.

Contribution of Authors

Study design: NSC, SKD, PS, SD and SR; data analysis, acquisition, and interpretation: NSC, SKD, SD and PS; quality assessment: PS; drafting of manuscript: NSC, SKD, PS, and SR.

Orcid

Nang S Choupoo https://orcid.org/0000-0001-6270-3981

Saurabh K Das https://orcid.org/0000-0001-7798-4528

Priyam Saikia https://orcid.org/0000-0001-6608-484X

Samarjit Dey https://orcid.org/0000-0001-8211-253X

Sumit Ray https://orcid.org/0000-0001-5192-4711