Expanding restrictive transfusion evidence in surgical practice: a multicentre, prospective cohort study

Abstract

Background

Findings of observational studies investigating the impact of transfusions are at odds with those of randomised controlled trials, raising concern that observational studies may be inappropriate to inform transfusion decisions. We examined whether observational data could replicate evidence from randomised controlled trials on restrictive transfusion in cardiac and orthopaedic surgery, and be generalised to broader specialties as well as to a lower haemoglobin transfusion threshold (7 g/dL).

Material and methods

A multicentre, prospective cohort study was performed at three representative regional hospitals in China between 2015 and 2016. Participants were surgical inpatients (≥18 years; hospital stay ≥24 h) in six specialties: cardiac, cerebral, vascular (CCV), and orthopaedic, general, thoracic (non-CCV). Patients with a stable haemoglobin (7–10 g/dL) constituted the primary analytic sample, while patients with ≥500 mL intra-operative bleeding were analysed separately to avoid haemoglobin instability. The association of transfusion with surgical outcomes (death, in-hospital complications) was evaluated.

Results

The transfusion rate was 10.7% in 36,607 patients (mean age, 52.5±14.3 years; 52.3% female). After restriction, stratification, and propensity score matching to reduce patients’ heterogeneity, transfusion was unrelated to death (CCV: odds ratio [OR]=0.74, 95% confidence interval [CI]: 0.16–3.39; non-CCV: OR 0.83, 95% CI: 0.36–1.94) and the composite complication (CCV: OR 1.31, 95% CI: 0.63–2.72; non-CCV: OR=1.24, 95% CI: 0.81–1.90). The results were consistent in subgroups (elderly, coronary heart disease, malignant tumour, severe illness) and applicable to patients with significant bleeding after restoration of a stable haemoglobin.

Discussion

Transfusion at a stable haemoglobin concentration of 7–10 g/dL did not alter surgical outcomes. Our results show the feasibility of observational data to expand restrictive transfusion to broader specialties and a lower transfusion threshold in surgical practice.

INTRODUCTION

More than half of red blood cell (RBC) transfusions are administered to surgical patients¹. As a result of the constantly increasing volume of surgical procedures and population aging, blood shortages are becoming a concerning challenge worldwide. Recent guidelines have recommended a restrictive transfusion policy for surgical patients^2–4, with support from 18 randomised controlled trials (RCT) that compared transfusion policies between low (restrictive) and high (liberal), based on haemoglobin concentrations (typically 8 vs 10 g/dL, and more recently, 7.5 vs 9 .5 g/dL), and showed no significant prognostic differences between these haemoglobin thresholds^2,5–6. However, these RCTs were primarily confined to cardiac or orthopaedic surgery^2,6,7. Thus, it remains uncertain whether the results can be generalised to non-cardiac and non-orthopaedic cases⁸. In addition, although the TRICC trial demonstrated the feasibility of an even lower haemoglobin threshold (7 g/dL) for critical care patients⁹, we aimed to provide further data about whether this threshold may also apply to patients undergoing surgery, which is a special set of patients characterised by greater heterogeneity.

There are a few planned and ongoing RCT in search of the minimum tolerable haemoglobin threshold^10,11; however, their speed is hampered by narrow patient and procedural spectra and examination of fixed thresholds⁸. Observational data are more easily obtainable, can be collected from larger sample sizes, and relate better to daily care¹². Observational studies therefore hold promise for more rapid translation and generalisation of evidence. The role of observational data in the evaluation of drug effects was highlighted in the 21^st Century Cures Act¹³. In the domain of transfusion research, however, most observational studies have shown increased mortality and morbidity after transfusion^14–19, findings which are at odds with those of RCT. In an attempt to explain such a long-existing systemic discrepancy, it is widely believed that observational studies cannot properly adjust for a “sicker patient” effect²⁰; in other words, transfused and non-transfused patients are deemed heterogeneous, and transfused patients are more physically unwell and are therefore prone to adverse outcomes²¹. An often overlooked, yet perhaps more critical distinction is that observational studies examine the outcomes of transfusion, whereas the intrinsic focus of transfusion RCT is the haemoglobin threshold²⁰.

By unifying the research objective of this observational study with that of RCT and tailoring the objective-related design and analysis, we studied whether observational data are useful for rapid evidence generation and translation in terms of restrictive transfusion. Specifically, we examined whether observational data could replicate existing RCT evidence in cardiac and orthopaedic surgery, be generalised to patients across extended surgical specialties, and be used to delineate a lower haemoglobin threshold (7 g/dL). Our main hypothesis was that transfusion at a stable haemoglobin concentration of 7–10 g/dL does not alter surgical outcomes across a wide variety of surgical specialties.

MATERIAL AND METHODS

Data sources

We used data from a multicentre prospective project on improving surgical safety²², which was conducted at four academic/teaching hospitals in China (project registered at: researchregistry5922). Ethical approval for this study was obtained from the institutional review board of Peking Union Medical College Hospital, Beijing, China (approval n.: S-574) on July 26 2013, and the requirement for written informed consent was waived because individual information was analysed anonymously. The study sites were chosen to represent the geographic diversity of China. One site located in the Qinghai-Tibetan Plateau was excluded, because it mainly treats residents who lived at a high altitude of 2,000–5,000 meters and for whom the common transfusion threshold in plain areas may not be applicable. The data collection periods ran in parallel (January to June 2015 and January to June 2016) according to the overall project design²². The transfusion pattern within these periods was unaffected, because transfusion practice was intended to be observational.

Study participants

To obtain a natural transfusion practice pattern, we first defined the base population as follows: age ≥18 years, hospital stay ≥24 hours, and surgery for any of six specialties in which RBC transfusions occur (cardiac, cerebral, vascular, orthopaedic, general, and thoracic surgery). To evaluate transfusion triggers in patients with a stable haemoglobin concentration, we applied the following conditions and derived a primary analytic sample: (i) intra-operative bleeding <500 mL (estimated by gauze weight and aspiration); (ii) a stable haemoglobin concentration of 7–10 g/dL, which was used to cover both the lower (7 g/dL) and classic (10 g/dL) thresholds. The haemoglobin concentration of interest was defined based on the last measurement before the initial RBC transfusion in transfused patients and the nadir during hospitalisation for non-transfused patients. An analysis of patients with an intra-operative blood loss volume of ≥500 mL was performed separately, because transfusion strategy may be different for patients with significant bleeding. A patient flowchart is presented in the Online Supplementary Content (Figure S1).

Exposure and outcomes

The study exposure was allogeneic RBC transfusion. Multiple transfusions were defined as receipt of ≥2 units of RBC unless they were within 24 hours. The study outcomes were death (in-hospital death or death within 30 days of discharge) and in-hospital complications. For comparability, the complication items were similar to those used in RCT^7,23, including ischaemic complications (myocardial infarction, stroke, acute renal failure); infection (surgical site infection, pneumonia, sepsis, septic shock, urinary tract infection); and other complications (cardiac arrest requiring cardiopulmonary resuscitation, heart failure, reintubation, mechanical ventilation for ≥48 hours post-operatively, atelectasis, respiratory failure, wound dehiscence, delayed incision healing, pulmonary embolism, venous thrombosis, and multiple organ dysfunction syndrome). The definitions of these complications are detailed in a previous report²⁴.

Covariates

Covariates included demographic characteristics (age, sex); body mass index; comorbidities in medical records or diagnosed during hospitalisation (coronary heart disease [CHD], diabetes mellitus, hypertension); American Society of Anesthesiologists (ASA) score; pre-operative laboratory parameters (albumin, creatinine, white blood cell count); and operative characteristics (emergency/elective surgery, surgical specialty, operating time, intra-operative volume of bleeding).

Data collection

Haemoglobin measurements and transfusion information were obtained directly from laboratory information systems and clinical blood bank systems, respectively, which captured all transfusion-related activities for each patient. Clinical and outcome information were collected via electronic data capture systems by a specialty-specific nurse and an attending surgeon, respectively. Outcome information was regularly monitored by a statistical team and reviewed against medical records by a clinical specialist team where necessary. Other data quality control measures detailed in the study protocol²² were followed closely.

Statistical analysis

To identify between-group differences in patient and operative characteristics, we used the χ² test or Fisher’s exact test for categorical variables and the t-test or Wilcoxon’s rank-sum test for continuous variables, as appropriate.

We progressively performed the following analyses to evaluate the safety of a lower haemoglobin threshold in surgical patients. First, in the primary analytic sample, outcomes were compared between transfused and non-transfused patients with a haemoglobin concentration of 7–10 g/dL. Comparison strategies used in previous RCT and in this study are presented in the Online Supplementary Content (Figure S2). To generalise the spectrum of patients, we stratified the sample into a CCV group (extended from cardiac surgery to include cerebral and vascular surgery) and a non-CCV group (extended from orthopaedic surgery to include general and thoracic surgery). These two groups have distinct characteristics in terms of transfusion and outcomes¹⁷. Second, similar comparisons were made in several subgroups of patients of particular clinical concern, including those with CHD, aged ≥65 years, with severe illness (ASA score ≥3) or malignant tumours. Third, outcomes following transfusion at lower (7–7.9 g/dL) and higher (8–10 g/dL) haemoglobin concentrations were compared to further verify the safety of the lower haemoglobin threshold. Finally, the applicability of the findings was assessed among patients with an intra-operative bleeding volume of ≥500 mL after obtaining relative haemoglobin-stability after three post-operative days. Logistic regression with adjustment for potential confounders was used in all analyses.

In all analyses, demographic and clinical variables that significantly affected both transfusion and patients’ outcomes were identified as key covariates and used in propensity score matching to minimise confounding effects. Nearest-neighbour caliper matching (1:1) was used with a caliper equal to 0.2 standard deviations of the propensity score. Balance in the covariate distribution between groups was assessed by probability distributions of propensity scores and quantified using standardised bias, for which ≤10% was considered acceptable²⁵.

Throughout the analyses, imputation of missing data was not attempted because of the low rate of missing data (0–9%; mostly 0–4%), which was comparable between the transfused and non-transfused groups, and the intention to retain the authenticity of the clinical data.

The data analysis and statistical plans were written after the data were accessed. All analyses were performed using SAS software, version 9.4 (SAS Institute, Cary, NC, USA) and R, version 3.6.2 (the R Foundation for Statistical Computing, Vienna, Austria). A two-tailed p value of <0.05 was considered statistically significant.

RESULTS

Base population

Natural transfusion pattern

Of 36,607 patients in the base population (mean age, 52.5 years; 52.3% female), 3,930 patients (10.7%) underwent RBC transfusion with a median of 3 units (interquatile range [IQR] 2–4 units). Intra-operative blood loss and anaemia were the primary indications for transfusion. The transfusion rate increased by 12 times for patients with an intra-operative bleeding volume of ≥500 mL compared with <200 mL (53.5 vs 4.1%) and by approximately seven times for patients with a haemoglobin concentration of <7 g/dL compared with 9–10 g/dL (57.3 vs 7.3%) (both p<0.01).

Patients’ characteristics and outcomes

As expected, substantial heterogeneity was found between patients who did and who did not undergo transfusion (Table I). Compared with non-transfused patients, transfused patients were more likely to be older (≥65 years: 28.0 vs 18.9%), have CHD (10.2 vs 5.6%), and have an ASA score of ≥3 (49.3 vs 21.8%). Transfused patients also had a lower haemoglobin concentration throughout hospitalisation (e.g., pre-operative mean: 12.0 vs 13.6 g/dL) and a longer operative time (median: 3.8 vs 1.9 h) (both p<0.05). Transfused patients showed significantly higher rates of death (2.2 vs 0.3%) and complications (e.g., composite: 10.0 vs 2.6%) compared with non-transfused patients (both p<0.01).

Table I.

Patients’ characteristics and outcomes stratified by transfusion status in the base population

	Transfusion (n=3,930)	Non-transfusion (n=32,677)	p
Characteristics
Male, n (%)	2,023 (51.5)	15,423 (47.2)	<0.001
Age ≥65 years, n (%)	1,099 (28.0)	6,163 (18.9)	<0.001
BMI, kg/m2, mean ± SD	23.3±3.6	24.0±3.5	<0.001
Comorbidity, n (%)
Coronary heart disease	400 (10.2)	1,841 (5.6)	<0.001
Diabetes mellitus	358 (9.1)	2,651 (8.1)	0.032
Hypertension	867 (22.1)	6,465 (19.8)	<0.001
Malignant tumour, n (%)	1,396 (35.5)	13,019 (39.8)	<0.001
ASA score ≥3, n (%)	1,924 (49.3)	6,923 (21.8)	<0.001
Pre-operative laboratory test, mean ± SD
Albumin, g/L	38.2±7.9	41.9±12.0	<0.001
Creatinine, μmol/L	78.5±53.4	72.1±37.7	<0.001
WBC count, ×109/L	7.4±5.3	6.8±3.3	<0.001
Haemoglobin level, g/dL, mean ± SD
Pre-operative	12.0±2.5	13.6±1.8	<0.001
Post-operative	10.5±2.1	12.4±1.9	<0.001
At discharge	10.2±1.9	12.0±1.9	<0.001
Operative characteristics
Emergency surgery, n (%)	340 (8.7)	1,666 (5.3)	<0.001
Surgery group, n (%)
CCV	1,245 (31.7)	8,061 (24.7)
non-CCV	2,685 (68.3)	24,616 (75.3)
Operating time, h, median (IQR)	3.8 (2.6 to 5.1)	1.9 (1.2 to 2.9)	<0.001
Intra-operative bleeding, ×100 mL, mean ± SD	6.2±6.2	1.7±1.7	<0.001
Outcomes, n (%)
Death	86 (2.2)	90 (0.3)	<0.001
Composite complications	394 (10.0)	838 (2.6)	<0.001
Ischaemic complications	87 (2.2)	95 (0.3)	<0.001
Infection	214 (5.5)	475 (1.5)	<0.001

BMI: body mass index; SD: standard deviation; ASA: American Society of Anesthesiologists; WBC: white blood cell; CCV: cardiac, cerebral, and vascular surgery; non-CCV: orthopaedic, general, and thoracic surgery; IQR: interquartile range

Primary analytic sample

Reduced patient heterogeneity

Of the 4,114 patients in the primary analytic sample, 125 of 1,101 patients in the CCV group (11.4%) underwent transfusion (initial transfusion in pre-operative, intra-operative, and post-operative stages: 12.8%, 26.4%, and 60.8%, respectively), while 495 of 3,013 patients in the non-CCV group (16.4%) underwent transfusion (initial transfusion in pre-operative, intra-operative, and post-operative stages: 21.4%, 35.4%, and 43.2%, respectively). The median volume of RBC given to transfused patients was 2 units (IQR 2–4 units). When stratified by study site, the transfusion rate and median transfusion volume were 17.9% and 2 units (IQR 2–4 units), respectively, in site A; 15.2% and 4 units (IQR 2–4 units) respectively, in site B; and 12.9% and 2 units (IQR 2–4 units), respectively, in site C. Notably, the imbalanced distribution of covariates was markedly reduced between transfused and non-transfused patients, although this imbalanced distribution remained from the base population to the primary analytic sample. Specifically, the difference between transfused and non-transfused patients in the rate of an ASA score ≥3 decreased from 2.3 times (49.3 vs 2 1.8%) to 1.1 times (74.0 vs 6 4.4%) in the CCV group and to 1.2 times (39.3 vs 32.5%) in the non-CCV groups.

After propensity score matching, 124 (matching rate: 99.2%) and 495 (matching rate: 100.0%) transfused patients were matched with non-transfused patients in the CCV and non-CCV groups, respectively, and between-group heterogeneity further decreased (Table II). Key covariates were well balanced. Specifically, the standardised bias for an ASA score of ≥3 decreased from 21.3% to 3.9% in the CCV group and from 14.5 to 0.9% in the non-CCV group, and the overall standardised bias decreased to ≤6.3%.

Table II.

Key covariates and patients’ outcomes before and after propensity score matching in the primary analytic sample

CCV group	Original				Matched
	Transfusion (n=125) n (%)	Non-transfusion (n=976) n (%)	p	Standardised bias, %	Transfusion (n=124) n (%)	Non-transfusion (n=124) n (%)	p	Standardised bias, %
Age ≥65 years	28 (22.4)	184 (18.9)	0.344	10.7	27 (21.8)	24 (19.4)	0.637	6.3
Male	57 (45.6)	342 (35.0)	0.021	22.7	57 (46.0)	55 (44.4)	0.799	1.7
ASA score ≥3	91 (74.0)	604 (64.4)	0.035	21.3	90 (73.8)	93 (76.2)	0.657	3.9
Haemoglobin*, g/dL			<0.001	74.5			0.799	1.7
7–7.9	59 (47.2)	153 (15.7)			58 (46.8)	56 (45.2)
8–10	66 (52.8)	823 (84.3)			66 (53.2)	68 (54.8)
Operating time ≥3 h	83 (69.2)	539 (60.7)	0.073	17.3	82 (68.9)	80 (67.8)	0.854	1.8
Study site			<0.001	36.7			0.876	2.6
Site A	45 (36.0)	195 (20.0)			44 (35.5)	45 (36.3)
Site B	22 (17.6)	179 (18.3)			22 (17.7)	19 (15.3)
Site C	58 (46.4)	602 (61.7)			58 (46.8)	60 (48.4)
Outcomes
Death	3 (2.4)	13 (1.3)	0.413		3 (2.4)	4 (3.2)	0.999
OR (95% CI)	1.82 (0.51 to 6.48)	1.0 (ref)			0.74 (0.16 to 3.39)	1.0 (ref)
Composite complications	19 (15.2)	68 (7.0)	0.001		19 (15.3)	15 (12.1)	0.460
OR (95% CI)	2.39 (1.39 to 4.14)	1.0 (ref)			1.31 (0.63 to 2.72)	1.0 (ref)
Ischaemic complications	6 (4.8)	17 (1.7)	0.038		6 (4.8)	3 (2.4)	0.500
OR (95% CI)	2.84 (1.10 to 7.35)	1.0 (ref)			2.05 (0.50 to 8.39)	1.0 (ref)
Infection	9 (7.2)	37 (3.8)	0.073		9 (7.3)	6 (4,8)	0.424
OR (95% CI)	1.97 (0.93 to 4.18)	1.0 (ref)			1.54 (0.53 to 4.46)	1.0 (ref)
Non-CCV group	Original				Matched
	Transfusion (n=495) n (%)	Non-transfusion (n=2,518) n (%)	p	Standardized bias,%	Transfusion (n=495) n (%)	Non-transfusion (n=495) n (%)	P	Standardized bias,%
Age ≥65 years	223 (45.1)	880 (35.0)	<0.001	21.5	223 (45.1)	222 (44.9)	0.949	0.4
Male	230 (46.5)	865 (34.4)	<0.001	23.4	230 (46.5)	226 (45.7)	0.799	0.4
ASA score ≥3	191 (39.3)	796 (32.5)	0.004	14.5	191 (39.3)	190 (39.8)	0.866	0.9
Haemoglobin*, g/dL			<0.001	55.0			0.597	0.4
7–7.9	183 (37.0)	344 (13.7)			183 (37.0)	175 (35.4)
8–10	312 (63.0)	2174 (86.3)			312 (63.0)	320 (64.7)
Operating time ≥3h	224 (47.8)	781 (33.0)	<0.001	30.5	224 (47.8)	223 (47.7)	0.973	<0.1
Study site			0.360	9.2			0.376	3.2
Site A	180 (36.4)	836 (33.2)			180 (36.4)	179 (36.2)
Site B	150 (30.3)	778 (30.9)			150 (30.3)	133 (26.9)
Site C	165 (33.3)	904 (35.9)			165 (33.3)	183 (37.0)
Outcomes
Death	10 (2.0)	19 (0.8)	0.019		10 (2.0)	12 (2.4)	0.666
OR (95% CI)	2.71 (1.25 to 5.87)	1.0 (ref)			0.83 (0.36 to 1.94)	1.0 (ref)
Composite complications	51 (10.3)	140 (5.6)	<0.001		51 (10.3)	42 (8.5)	0.327
OR (95% CI)	1.95 (1.39 to 2.73)	1.0 (ref)			1.24 (0.81 to 1.90)	1.0 (ref)
Ischaemic complications	9 (1.8)	20 (0.8)	0.043		9 (1.8)	11 (2.2)	0.651
OR (95% CI)	2.31 (1.05 to 5.11)	1.0 (ref)			0.81 (0.33 to 1.98)	1.0 (ref)
Infection	35 (7.1)	84 (3.3)	<0.001		35 (7.1)	21 (4.2)	0.054
OR (95% CI)	2.20 (1.47 to 3.31)	1.0 (ref)			1.72 (0.98 to 2.99)	1.0 (ref)

All the covariates in the table were included in the propensity score calculation.

^*Haemoglobin concentration was defined as the last measurement before the initial red blood cell transfusion for transfused patients and the nadir during hospitalisation for non-transfused patients.

ASA, American Society of Anesthesiologists; CCV, cardiac, cerebral, and vascular surgery; non-CCV, orthopaedic, general, and thoracic surgery; OR, odds ratio; CI, confidence interval; ref, reference.

Outcome comparison

Figure 1 shows the impact of the gradual reduction in patients’ heterogeneity on the results of the outcome comparison. From the base population to the primary analytic sample, the unadjusted odds ratios (OR) of death and composite complications decreased from 10.24 to 2.32 and from 4.73 to 2.01, respectively. After propensity score matching (Table II), transfusion was no longer related to the risk of death (CCV: OR=0.74, 95% confidence interval [CI]: 0.16–3.39; non-CCV: OR=0.83, 95% CI: 0.36–1.94) or composite complications (CCV: OR=1.31, 95% CI: 0.63–2.72; non-CCV: OR=1.24, 95% CI: 0.81–1.90). Considering the variability in transfusion and outcomes among different study sites, propensity score matching was conducted at each site (Online Supplementary Content, Table SIa, Ib, and Ic). Overall, the results remained consistent with the primary findings, except that transfusion showed an increased risk of composite complications at site B. The specialty-specific results are shown in the Online Supplementary Content, Table SII. For instance, transfusion was unrelated to death and all complications in cardiac surgery (all p>0.05 after adjustment) and unrelated to death and complications (except for infection, OR=3.13, 95% CI: 1.20–8.19) in orthopaedic surgery.

Figure 1.

Effect of reduced patients’ heterogeneity on transfusion effect estimation

*Haemoglobin concentration was defined as the last measurement before the initial red blood cell transfusion for transfused patients and the nadir during hospitalisation for non-transfused patients. Hb, haemoglobin; CCV, cardiac, cerebral, and vascular surgery; non-CCV, orthopaedic, general, and thoracic surgery; OR, odds ratio; CI, confidence interval; RBC, red blood cell.

Subgroup analysis

We performed a comparison similar to the primary analysis in subgroups of patients who were more likely to undergo transfusion (Figure 2). Transfusion and non-transfusion at a haemoglobin concentration of 7–10 g/dL were not associated with the study outcomes in all subgroups after matching. For instance, among patients with CHD, the OR was 1.34 (95% CI: 0.56–3.17) for composite complications. Similarly, the corresponding OR was 1.32 (95% CI: 0.63–2.73) in patients with malignant tumours.

Figure 2.

Outcome comparison of transfusion vs non-transfusion at a haemoglobin concentration of 7–10 g/dL in several subgroups of patients

*The key covariates included in the calculation of propensity score matching were age, sex, ASA score, haemoglobin concentration, operating time, surgery group (CCV and non-CCV), and study site. OR: odds ratio; CI: confidence interval; ASA: American Society of Anesthesiologists; CCV: cardiac, cerebral, and vascular surgery; non-CCV: orthopaedic, general, and thoracic surgery.

Patients transfused at different thresholds

Almost all key covariates showed naturally balanced distributions in patients transfused at a haemoglobin concentration of 7–7.9 g/dL and those transfused at a haemoglobin concentration of 8–10 g/dL (Table III). After adjustment, there was no significant increase in outcome risk in the lower vs higher haemoglobin concentration group (death: OR=0.97, 95% CI: 0.31–3.02; composite complications: OR=0.69, 95% CI: 0.41–1.19). To enhance the validity of this comparison, we further examined the haemoglobin concentration measurement frequency (Online Supplementary Content, Table SIII) and the mean daily haemoglobin concentration in the lower and higher haemoglobin concentration groups (Online Supplementary Content, Figure S3). Transfused patients underwent significantly more tests compared with non-transfused patients (median: 7 vs 4, IQR: 5–9 vs 3–7). Haemoglobin concentration increased markedly within 24 hours after initial transfusion in both groups and remained steady thereafter. The mean ± standard error of haemoglobin concentration in the lower and higher threshold groups was 7.4±0.2 g/dL and 8.8±0.3 g/dL before transfusion, respectively, and differed by approximately 0.5 g/dL after transfusion.

Table III.

Comparison of transfused patients in the two threshold groups in the primary analytic sample

	Transfused at heamoglobin level of		p
	7–7.9 g/dL (n=242) n (%)	8–10 g/dL (n=378) n (%)
Age ≥65 years	96 (39.7)	155 (41.0)	0.741
Male	111 (45.9)	176 (46.6)	0.866
ASA score ≥3	113 (47.9)	169 (45.3)	0.535
Intra-operative blood loss volume, mL			0.972
< 200	120 (52.6)	190 (52.8)
200–500	108 (47.4)	170 (47.2)
Operating time ≥3 h	109 (47.8)	198 (54.9)	0.096
Surgery group			0.036
CCV	59 (24.4)	66 (17.5)
Non-CCV	183 (75.6)	312 (82.5)
Study site			0.041
Site A	76 (31.4)	149 (39.4)
Site B	65 (26.9)	107 (28.3)
Site C	101 (41.7)	122 (32.3)
Outcomes
Death	5 (2.1)	8 (2.1)	0.966
OR (95% CI)	0.97 (0.31 to 3.02)	1.0 (ref)
Composite complications	22 (9.1)	48 (12.7)	0.166
OR (95% CI)	0.69 (0.41 to 1.19)	1.0 (ref)
Ischaemic complications	8 (3.3)	7 (1.9)	0.250
OR (95% CI)	1.82 (0.64 to 5.14)	1.0 (ref)
Infection	12 (5.0)	32 (8.5)	0.097
OR (95% CI)	0.56 (0.28 to 1.11)	1.0 (ref)

Adjusted for surgery group and study site in calculating odds ratio (95% confidence interval). ASA, American Society of Anesthesiologists; CCV, cardiac, cerebral, and vascular surgery; non-CCV, orthopaedic, general, and thoracic surgery; OR, odds ratio; CI, confidence interval; ref, reference.

Patients with significant bleeding

We further assessed whether transfusion vs non-transfusion at a haemoglobin concentration of 7–10 g/dL was related to outcomes in patients with an intra-operative bleeding volume of ≥500 mL (Table IV). To avoid the impact of intra-operative bleeding on the decision to perform transfusion, haemoglobin concentration was defined as the last reading within 3 days after surgery (when the haemoglobin concentration of nearly all patients had become stable), and transfusion was defined after that time. The results were very similar to those of the primary analytic sample. Specifically, transfused and non-transfused patients were heterogeneous (i.e., the standardised bias was 9.5–18.1% regarding age, sex, and ASA score); Moreover, transfused patients demonstrated higher mortality (2.7 vs 0.3%) and morbidity (composite complication: 14.0 vs 8.2%) rates. After propensity score matching, between-group patients’ heterogeneity diminished (i.e., the standardised bias was <0.1% regarding age, sex and ASA score), and the results indicated equal patients’ outcomes (all p>0.05 for death and complications).

Table IV.

Key covariates and patients’ outcomes before and after propensity score matching in patients with an intra-operative bleeding volume ≥500 mL^*

	Original				Matched
	Transfusion (n=150)	Non-transfusion (n=967)	p	Standardised bias, %	Transfusion (n=149)	Non-transfusion (n=149)	p	Standardised bias, %
	n (%)	n (%)			n (%)	n (%)
Age ≥65 years	42 (28.0)	196 (20.3)	0.031	18.1	41 (27.5)	41 (27.5)	0.999	<0.1
Male	84 (56.0)	464 (48.0)	0.068	16.2	83 (55.7)	83 (55.7)	0.999	<0.1
ASA score ≥3	98 (65.3)	587 (60.8)	0.285	9.5	97 (65.1)	97 (65.1)	0.999	<0.1
Intra-operative blood loss, mL			0.241	11.4			0.956	3.4
500−	95 (63.3)	645 (66.7)			95 (63.8)	97 (65.1)
1000−	24 (16.0)	184 (18.0)			24 (16.1)	24 (16.1)
1500−	31 (20.7)	148 (15.3)			30 (20.1)	28 (18.8)
Study site			0.056	20.4			0.931	4.4
Site A	53 (35.3)	251 (26.0)			52 (34.9)	54 (36.2)
Site B	20 (13.3)	143 (14.8)			20 (13.4)	18 (12.1)
Site C	77 (51.3)	573 (59.3)			77 (51.7)	77 (51.7)
Outcomes
Death	4 (2.7)	3 (0.3)	0.008		4 (2.7)	2 (1.3)	0.684
OR (95% CI)	8.80 (1.95 to 39.73)	1.0 (ref)			2.03 (0.37 to 11.24)	1.0 (ref)
Composite complications	21 (14.0)	79 (8.2)	0.020		20 (13.4)	15 (10.1)	0.368
OR (95% CI)	1.83 (1.09 to 3.06)	1.0 (ref)			1.39 (0.68 to 2.82)	1.0 (ref)
Ischaemic complications	7 (4.7)	14 (1.5)	0.015		7 (4.7)	6 (4.0)	0.777
OR (95% CI)	3.33 (1.32 to 8.40)	1.0 (ref)			1.17 (0.39 to 3.58)	1.0 (ref)
Infection	9 (6.0)	42 (4.3)	0.366		9 (6.0)	4 (2.7)	0.156
OR (95% CI)	1.41 (0.67 to 2.95)	1.0 (ref)			2.33 (0.70 to 7.74)	1.0 (ref)

^*Comparisons were made between patients who underwent transfusion and those who did not, indicated by a haemoglobin concentration of 7–10 g/dL (defined as the last measurement within 3 post-operative days). All the covariates in the table were included in the propensity score calculation.

ASA, American Society of Anesthesiologists; OR, odds ratio; CI, confidence interval; ref, reference.

DISCUSSION

Our analytic results show that transfusion in patients with a stable haemoglobin concentration of 7–10 g/dL is not associated with death, ischaemic complications, and other complications. These results are in line with existing evidence on restrictive transfusion practices, but the present results were derived from more extensive surgical specialties and using a lower haemoglobin concentration threshold of 7 g/dL, which was consistent in subgroups of patients with a higher likelihood of transfusion and those who experienced significant bleeding. This suggests a substantially lower RBC requirement; indeed, according to a rough estimate based on the transfusion pattern in this study and national data²⁶, the overall RBC demand in China could be reduced by 0.47 million units annually if a transfusion threshold of 7 g/dL rather than 8 g/dL was applied across these six surgical specialties.

As shown in Figure 1, all the approaches applied in the design and analysis of this observational study (restriction, stratification, and propensity score matching) were intended to reduce patients’ heterogeneity, which is critical for valid effect estimation using real-world data²⁷. In previous observational studies and RCT it was common practice to restrict observations to a specific disease or surgical specialty^28–30. Nevertheless, the heterogeneity in transfusion effect estimation in observational studies was very much larger than in RCT (I² = 96.0 vs 0.0% for mortality in cardiac surgery) according to a meta-analysis of 64 studies¹⁸. Such a sharp contrast indicates the need to reflect on how best to define a homogeneous study population. In this study, the base population was highly heterogeneous and the results of outcome comparisons were similar to those of most previous observational studies (transfusion was associated with an increased risk of adverse surgical outcomes). The primary analysis was limited to patients with a stable haemoglobin concentration of 7–10 g/dL, to unify with the objectives of transfusion RCT (transfusion threshold)^31,32. This change from a specialty-based to a haemoglobin concentration-based definition of the study population has methodological relevance, Specifically, the new definition (i) naturally excludes severe anaemia and irrational transfusion at higher concentrations³³, greatly reducing between-group patients’ heterogeneity in key covariates (note the differences in Tables I and II), and (ii) eases concerns regarding specialty-specific differences (given that the results of previous RCT are largely consistent)^7,24,33. Thus, using this approach would be helpful to accelerate evidence generation and translation beyond the boundary of a specific specialty.

A systematic review of 17 RCT has shown that transfusion increases infection risk in patients undergoing hip or knee replacement surgery, but not cardiac surgery³⁴. Our results are similar to these findings, even after extending from orthopaedic surgery to the non-CCV group and from cardiac surgery to the CCV group, indicating that these extensions are valid. In terms of infectious complications, our results show that the effect of transfusion in the non-CCV group is close to statistical significance (OR = 1.72, 95% CI: 0.98–2.99). One of the possible mechanisms of infection after transfusion is transfusion-induced immunomodulation, which results in transfusion-related lung injury or immunosuppression and increases the subject’s susceptibility to infection³⁵. However, the reasons for differences among surgical specialties require further investigation.

Our results, including the primary analysis, subgroup analysis, and significant bleeding analysis, are supportive of the lower haemoglobin concentration threshold of 7 g/dL that has been confirmed in RCT in critical care. These findings were intensified by our frequent haemoglobin concentration measurements and the stability of haemoglobin concentration in the two threshold groups, and the results are comparable with those of RCT^23,32. Despite the subgroups in this study showing a marked tendency toward transfusion, the optimal transfusion thresholds remain uncertain. For instance, a meta-analysis in 2018 showed that restrictive transfusion could decrease RBC consumption by 24% among patients with cardiovascular disease. Although no adverse effects on short-term mortality or morbidity were observed, separate results for patients with CHD are not available³⁶. For patients with malignant tumours, the tumour types were primarily intestinal and gastric tumours (38.2% and 17.6%, respectively). We identified only one RCT that included surgical oncology patients requiring post-operative intensive care⁶. Although the tumour types were similar to those in the present study, patients were not comparable in terms of the clinical setting. Our study provides preliminary results in these subgroups; however, we suggest further validation to establish well-grounded recommendations.

In propensity score matching analysis, comparisons of matched and unmatched patients are crucial for interpretation of the results³⁷, considering the possibility of bias caused by excluding large proportions of unmatched patients (CCV: 87.3%; non-CCV: 80.3%). We found that both patients’ characteristics and outcomes were significantly different between matched and unmatched non-transfused patients (Online Supplementary Content, Table SIV), which is strongly suggestive of a “sicker patient” effect. Of note, unknown or unmeasured confounding factors are always present in observational data and can be balanced only with randomisation. Thus, we do not suggest that transfusion RCT can be replaced, but the respective roles of RCT and observational studies should be better harnessed¹⁴. Only if these two types of study are co-ordinated and used to complement each other, can the overall body of evidence in transfusion medicine move forward in an efficient and reliable way.

One major contribution of this study is the generalisation of existing RCT evidence to support the safety of a lower haemoglobin threshold of 7 g/dL in critical care, expanding this to broader surgical specialties. More importantly, we reconsidered the primary source that led to systematic bias in observational transfusion studies, i.e., marked heterogeneity of patients, and addressed this issue through both design of the study and analysis of the findings. As such, we provide a study paradigm to use widely available observational data to accelerate the translation of restrictive transfusion evidence from RCT into surgical practice, paving a way for further studies.

Several limitations of the present study are worth noting. First, in the primary analysis sample, because each transfused patient was matched with a non-transfused counterpart in the 1:1 propensity score matching analysis and only a limited proportion of patients underwent RBC transfusion (15.1%), the sample size of the matching cohort in both the CCV and non-CCV groups was relatively small, which led to a lower power; Thus, some weak effects of transfusion may not have been identified. The relationship of transfusion to infectious complications in the non-CCV group in this analysis warrants further study to clarify the finding. Second, data availability limited our analyses to in-hospital outcomes, which may have contributed to the lower incidence of complications. Also, the long term effects of transfusion could not be assessed in this study. Third, for patients who underwent an initial RBC transfusion intra-operatively, the haemoglobin concentration considered was most frequently the last pre-operative measurement, because haemoglobin concentration is rarely measured intra-operatively. However, the haemoglobin concentration was not likely to be markedly influenced because we excluded patients with significant bleeding from the primary analytic sample. Fourth, for patients undergoing multiple transfusions (16% of transfused patients), we considered only the initial pre-transfusion haemoglobin concentration. Finally, our subgroup analysis was preliminary, and CHD was not classified into acute vs chronic forms. Findings of specialised RCT are expected to verify the validity of our results³⁸.

CONCLUSIONS

Our results suggest that restrictive transfusion is likely to be feasible in a more diverse surgical population than is currently included in the spectrum of patients studied in existing RCT and at a lower haemoglobin concentration threshold of 7 g/dL. The study shows that, by harmonising observational study objectives with those of RCT using tailored designs and analyses, observational studies have great potential in translating and augmenting restrictive transfusion evidence into surgical practice.