Patterns of neuropsychological changes after general anaesthesia in young children: secondary analysis of the Mayo Anesthesia Safety in Kids study

Abstract

Background

We hypothesised that exposure to multiple, but not single, procedures requiring general anaesthesia before age 3 yr is associated with a specific pattern of deficits in processing speed and fine motor skills.

Methods

A secondary analysis (using factor and cluster analyses) of data from the Mayo Anesthesia Safety in Kids study was conducted, in which unexposed, singly exposed, and multiply exposed children born in Olmsted County, MN, USA from 1994 to 2007 were sampled using a propensity-guided approach and underwent neuropsychological testing at ages 8–12 or 15–20 yr.

Results

In the factor analysis, the data were well fit to a five factor model. For subjects multiply (but not singly) exposed to anaesthesia, a factor reflecting motor skills, visual-motor integration, and processing speed was significantly lower [standardised difference of –0.35 (95% confidence interval {CI} –0.57 to –0.13)] compared with unexposed subjects. No other factor was associated with exposure. Three groups were identified in the cluster analysis, with 106 subjects (10.6%) in Cluster A (lowest performance in most tests), 557 (55.9%) in Cluster B, and 334 (33.5%) in Cluster C (highest performance in most tests). The odds of multiply exposed children belonging to Cluster A was 2.83 (95% CI: 1.49–5.35; P=0.001) compared with belonging to Cluster B; there was no other significant association between exposure status and cluster membership.

Conclusions

Multiple, but not single, exposures to procedures requiring general anaesthesia before age 3 yr are associated with a specific pattern of deficits in neuropsychological tests. Factors predicting which children develop the most pronounced deficits remain unknown.

Editor's key points.

• •Based on deficits in learning and behaviour observed after exposure of young animals to general anaesthesia, a critical question is whether such effects also occur in children.
• •Based on secondary outcomes observed in the Mayo Anesthesia Safety in Kids (MASK) study, the authors hypothesised that exposures to multiple, but not single, procedures requiring general anaesthesia are associated with neurobehavioural deficits.
• •In factor and cluster analyses of the MASK study data, multiple, but not single, exposures to general anaesthesia before age 3 yr were associated with specific deficits in neuropsychological tests.
• •The specific pattern of differences argues against confounding by indication, but these analyses cannot directly demonstrate causality.

Several studies have sought evidence that the deficits in learning and behaviour caused by the exposure of young animals to general anaesthesia1, 2, 3, 4, 5, 6 also occur in children. Given the wide range of study designs and outcomes used, it is perhaps not surprising that the results have varied.7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 The objective of the Mayo Anesthesia Safety in Kids (MASK) study was to examine the association between exposure to procedures requiring general anaesthesia before the third birthday and neuropsychological function.²³ Prospective neuropsychological testing of a propensity-guided sample of a birth cohort of children was conducted, examining children unexposed, singly exposed, and multiply exposed to procedures requiring general anaesthesia before age 3 yr. The pre-specified outcome for the primary analysis was the full-scale intelligence-quotient score of the Wechsler Abbreviated Scale of Intelligence (WASI). An initial report described this primary analysis, finding that anaesthesia exposure was not associated with differences in this outcome.²⁴ Multiple tests of several other domains of neuropsychological function were also reported as secondary outcomes. Processing speed and fine motor abilities were decreased in multiply, but not singly, exposed children; other domains, including attention, problem solving, and memory, were not different amongst exposure groups. Although the pattern of exposure-related differences was relatively consistent within affected domains, individual tests were analysed independently and not corrected for multiple comparisons. For this reason, conclusions based on these secondary outcomes were tentative and require evaluation using techniques better suited for the multiple tests administered. Also, the analysis evaluated differences in means for each test derived from the subjects tested. If only some children in the population are affected, this may limit the ability to detect effects in these children.²⁵

Based on the pattern of secondary outcomes observed in the initial report of the MASK study,²⁴ we hypothesised that exposure to multiple, but not single, procedures requiring general anaesthesia would be associated with a specific pattern of deficits in processing speed and fine motor skills that may have consequences for subsequent learning and behaviour in some children. The aim of this secondary analysis of data from the MASK study was to evaluate this hypothesis utilising two analytical methods. The factor analysis sought to identify underlying latent variables (factors) measured by the tests based on variations in the individual tests across the subjects, thus reducing these many observations into a fewer number of dimensions. We then determined the association between exposure status and these factors. The cluster analysis sought to group study participants, such that the pattern of testing results within a given group were more similar to each other than other groups, with groups denoted as ‘clusters’. We then determined whether membership in a cluster was associated with exposure status and parent reports of problems with learning and behaviour.

Methods

This study was approved by the Mayo Clinic and Olmsted Medical Center Institutional Review Boards, and written informed consent/assent was obtained. Study methods have been published previously.23, 24

Summary of recruitment and testing procedures

Individuals born from January 1, 1994 to December 31, 2007 in Olmsted County, MN, USA who lived within Olmsted County until their third birthday and who resided within 25 miles of Rochester, MN at the time of study according to available records at study onset were identified. Anaesthesia exposure status at 3 yr of age was classified as unexposed, singly, or multiply exposed through a review of medical records. Subjects were selected for recruitment using a frequency-matched approach, with strata defined based on their propensity for receiving single- and multiple-exposure general anaesthesia. Each subject was tested once, either when 8–12 or 15–20 yr old, to determine if any observed effects would change as children matured. Each subject was assessed by a trained psychometrist, generating 31 test scores for each subject (Supplementary file S1).23, 24 Parent/guardian questionnaires assessed perceived behaviour and learning difficulties (Supplementary file S2).

Analyses

Weighting procedures

Although propensity matching guided the recruitment strategy, not all who were invited agreed to participate, so that the relative frequencies of those with no, single, and multiple exposures to anaesthesia differed across sampling strata. Thus, analyses used inverse probability of treatment weighting (IPTW) to account for imbalances across exposure categories amongst children actually tested. Propensity scores were estimated using multinomial logistic regression with exposure status as the dependent variable.24, 26 Sex-by-characteristic and age-at-testing interactions were also included. Using these propensity scores, weights were calculated for each individual.

Factor analysis

Study subjects were randomly split into two half-samples, one for an initial exploratory factor analysis (EFA; n=499) and one for a confirmatory factor analysis (CFA; n=498), based on the factors identified in the initial EFA. Both EFA and CFA were conducted using the R (www.R-project.org) packages psych²⁷ and lavaan.²⁸

For the initial EFA, the number of factors was chosen based on inspection of scree plots and eigenvalues, with a preference for the most parsimonious solution that could be interpreted within the context of the underlying variables (i.e. a preference for the simplest solution). Factors were extracted using promax rotation, which allows for correlation between factors. The fit index of the model was evaluated using the standardised root-mean-square (RMS) residual.29, 30, 31 For the initial CFA, the model was fitted with factors identified from the EFA. Fit was evaluated using guidelines indicative of desirable fitting, including comparative fit index, standardised RMS residual, and RMS error of approximation.29, 30, 31 Fit was improved by freely estimating some residual correlations using the modification indices.

Once the factor structure was determined, factor scores derived from the EFA model were calculated using the Bartlett approach.³² Factor scores derived from the CFA model were calculated using the maximum likelihood estimation. For both EFA and CFA models, factor scores were calculated using all subjects. All factor scores were transformed to mean=0 and standard deviation (sd)=1. Linear regression, including IPTW weights, evaluated the relationship between exposure status and factor scores, using generalised estimating equations and a robust variance. Age at testing (8–12 vs 15–20 yr) was also included in the models. A two-tailed P-value of <0.05 was considered statistically significant for the overall two degree-of-freedom tests across exposure categories. Pairwise comparisons of single and multiple exposures vs no exposure were performed using P<0.025 (Bonferroni adjustment) to denote statistical significance.

Cluster analysis

The cluster analysis was performed on the entire study sample using scores derived from the administered neuropsychological tests (Supplementary file S1). This method can describe patient phenotypes without the need to make assumptions about how variables are related.33, 34, 35 Hierarchical agglomerative clustering using Ward's minimum variance method was used to cluster study participants based on test scores.³⁶ For ease of interpretation, it was decided a priori that between three and six clusters would be selected to facilitate interpretability of potential phenotypes and maintain adequate sample sizes in each cluster.

After clusters were determined, the mean and 95% confidence interval (CI) of the unstandardised and standardised (with each test score transformed to have a mean=0 and sd=1) outcome measures were descriptively summarised by cluster. Distributions of baseline characteristics and parent report variables across cluster membership were compared using analysis of variance F-tests for continuous variables and χ² or Fisher's exact tests (where appropriate) for categorical variables. Differences between the mean baseline characteristic within each cluster and the overall population norms were compared using one sample Z-tests. Logistic regression, including weighting with IPTW, evaluated the relationship between exposure status (categorised as 0, 1, and ≥2 exposures) and cluster membership, using generalised estimating equations and a robust variance. Age at testing (8–12 vs 15–20 yr) was also included in the models. Baseline characteristics were evaluated as potential moderators of the effect of exposure status on cluster membership. For each, regression analyses were performed, which included explanatory variables for exposure category, the potential moderator variable, and the moderator-by-exposure interaction effect. Finally, the association between each baseline characteristic and cluster membership was analysed using univariate multinomial logistic regression.

Analyses were performed using SAS (SAS Institute, Inc., Cary, NC, USA) version 9.4 and R for factor analyses.

Results

Factor analysis

In the initial EFA conducted on a split half-sample, six factors accounted for 57% of the variability, and five of these factors accounted for 53% of variability. Both five- and six-factor models demonstrated good fit; based on the principle of parsimony, a five-factor solution was selected (Supplementary file S3). The factor structure determined from the initial EFA was applied to the other split half-sample in a CFA model. Variables that had a factor loading of >0.30 were included in the model. One test (Wide Range Assessment of Memory and Learning, Second Edition [WRAML2] subtest: design recognition) did not have salient loadings on any factor in the EFA analysis and was excluded from the CFA model. Conners's Continuous Performance Test II (CPT-II) hit reaction time variable loaded on two factors (Factors 3 and 4), but is interpreted differently based on the positive or negative value. The fit indices of this initial CFA model indicated a relatively poor fit. Therefore, in the final CFA model, residuals were allowed to correlate based on modification indices to achieve a better fit (Table 1 and Fig 1). This resulted in a comparative fit index of 0.92, a standardised RMS residual of 0.069, and a RMS error of approximation of 0.057. (Supplementary file S4 presents the factor correlation and covariance matrix for the CFA.) Factor 1 appears to generally reflect motor skills, visual-motor integration, and processing speed measures. Factor 2 seems to best represent measures of verbal comprehension, especially verbal learning and memory. Factor 3 best reflects inattention, coupled with slow to variable reaction to target stimuli on CPT-II, whilst Factor 4 represents impulsivity on the same measure. Factor 5 represents a factor that reflects the ability to shift mentally in response to environmental stimuli (i.e. executive function).

Table 1.

Factor loadings for confirmatory factor analysis for all subjects (n=997). Values reflect the relationship of each variable to the underlying factor. They can be interpreted as standardised regression coefficients expressing the correlation between that variable and the underlying factor. CPT II, Conners's Continuous Performance Test II; CTOPP, Comprehensive Test of Phonological Processing; D-KEFS, Delis–Kaplan Executive Function System; WASI, Wechsler Abbreviated Scale of Intelligence; WRAML2, Wide Range Assessment of Memory and Learning, Second Edition.

Test	Factor 1	Factor 2	Factor 3	Factor 4	Factor 5
WASI: matrix reasoning subtest	0.73
WRAML2: attention/concentration index	0.87
D-KEFS trail making test: condition 1	0.77
D-KEFS trail making test: condition 2	0.83
D-KEFS trail making test: condition 3	1.00
D-KEFS trail making test: condition 4	0.96
D-KEFS trail making test: condition 5	0.76
CTOPP: rapid naming composite score	0.40
D-KEFS tower test: total achievement score	0.52
Grooved Pegboard dominant hand	0.79
Grooved Pegboard other hand	0.76
Beery: motor coordination	0.64
Beery: visual-motor integration	0.66
Beery: visual perception	0.58
WASI: vocabulary subtest		0.52
WRAML2: verbal memory index		1.00
WRAML2: story memory delay recall subtest		0.82
WRAML2: verbal learning delay recall subtest		0.75
WRAML2: story memory recognition subtest		0.58
WRAML2: verbal learning recognition subtest		0.52
WRAML2: design memory subtest		0.54
D-KEFS verbal fluency: category fluency		0.37
CPT-II: # omissions			0.54
CPT-II: # hit reaction time standard error			0.76
CPT-II: rariability			1.00
CPT-II: commissions				1.00
CPT-II: hit reaction time				–0.48
CPT-II: detectability				0.83
Wisconsin Card Sort: perseverative responses					1.00
Wisconsin Card Sort: perseverative errors					0.99

Fig 1.

Factor structure for the final confirmatory factor analysis model. Factors are depicted on the left (F1–F5), with scores loading on these factors in the middle. Double-headed arrows on the right indicate residual correlations. CPT II, Conners's Continuous Performance Test II; CTOPP, Comprehensive Test of Phonological Processing; D-KEFS, Delis–Kaplan Executive Function System; WASI, Wechsler Abbreviated Scale of Intelligence; WRAML2, Wide Range Assessment of Memory and Learning, Second Edition.

For factor scores calculated across all subjects using the CFA model, single exposures to anaesthesia were not significantly associated with differences in any factor compared with subjects unexposed to anaesthesia (Table 2). For subjects multiply exposed to anaesthesia, Factor 1 was significantly lower compared with unexposed subjects. No other factor differed significantly in multiply exposed children. The results were similar for factor scores calculated using the EFA model (Table 2).

Table 2.

Exposure status and factor scores derived from exploratory and confirmatory factor analyses. Values are the estimate (95% CI) of the difference from unexposed for all subjects (n=997). *P-values include the two degree-of-freedom P-value comparing across all exposure groups (significance indicated by P<0.05) and pairwise comparisons of single or multiple exposures vs no exposure; significance of pairwise comparisons indicated in bold for P<0.025 (Bonferroni correction). The analysis used inverse probability of treatment weighting (IPTW) to account for imbalances across exposure categories amongst children actually tested. A separate model was fit for each factor score. The reference group for anaesthesia exposure was no exposure. CFA, confirmatory factor analysis; CI, confidence interval; EFA, exploratory factor analysis.

Characteristic	Factor 1		Factor 2		Factor 3		Factor 4		Factor 5
	Mean (95% CI)	P-value*	Mean (95% CI)	P-value*	Mean (95% CI)	P-value*	Mean (95% CI)	P-value*	Mean (95% CI)	P-value*
EFA		<0.001		0.407		0.962		0.729		0.063
Unexposed	Reference		Reference		Reference		Reference		Reference
Single	–0.08 (–0.26, 0.10)	0.40	0.13 (–0.06, 0.31)	0.18	0.01 (–0.22, 0.25)	0.91	–0.02 (–0.23, 0.19)	0.87	0.10 (–0.08, 0.28)	0.29
Multiple	–0.41 (–0.62, –0.19)	<0.001	0.04 (–0.14, 0.23)	0.64	–0.02 (–0.23, 0.19)	0.86	–0.07 (–0.24, 0.11)	0.44	–0.14 (–0.35, 0.06)	0.16
CFA		0.005		0.187		0.882		0.834		0.032
Unexposed	Reference		Reference		Reference		Reference		Reference
Single	–0.05 (–0.23, 0.14)	0.63	0.09 (–0.09, 0.28)	0.33	–0.01 (–0.24, 0.23)	0.95	–0.00 (–0.22, 0.22)	0.99	0.12 (–0.06, 0.29)	0.20
Multiple	–0.35 (–0.57, –0.13)	0.002	–0.10 (–0.29, 0.10)	0.32	0.04 (–0.16, 0.24)	0.67	0.05 (–0.13, 0.23)	0.59	–0.14 (–0.34, 0.05)	0.14

Cluster analysis

Using cubic clustering criterion and semi-partial R-square, the optimum number of clusters was determined to be 3, with 106 subjects (10.6%) in Cluster A, 557 (55.9%) in Cluster B, and 334 (33.5%) in Cluster C. For each test, the mean values of the standardised differences for subjects in Cluster A indicated lower performance relative to the mean of all subjects, with the 95% CI for the mean not including 0 (Table 3). The mean values of the standardised differences in scores for subjects in Cluster C indicated higher performance relative to the mean, with the 95% CI not including 0, with the exception of the CPT-II hit reaction time. The mean values of the standardised differences for subjects in Cluster B were intermediate between those of Clusters A and C for all but three scores.

Table 3.

Values of psychometrist-administered neuropsychological variables according to cluster membership; n indicated number of subjects in a cluster. Values are means (95% confidence interval). ^∗Differences from the mean for all subjects tested, expressed as a fraction of the standard deviation (sd) for each test. Values in bold are ≥|1 sd|. Mean scores presented according to type of scale as follows: TS, T-score (mean=50; sd=10); SS, scaled score (mean=10; sd=3); StdS, standard score (mean=100; sd=15); ZS, Z-score (mean=0; sd=1). ^†Scores not included in the analysis used to determine clusters. CPT-II, Conners's Continuous Performance Test II; CTOPP, Comprehensive Test of Phonological Processing; D-KEFS, Delis–Kaplan Executive Function System; WASI, Wechsler Abbreviated Scale of Intelligence; WRAML2, Wide Range Assessment of Memory and Learning, Second Edition.

Characteristic	Mean scores			Differences from mean for all subjects*
	Cluster A (n=106)	Cluster B (n=557)	Cluster C (n=334)	Cluster A (n=106)	Cluster B (n=557)	Cluster C (n=334)
WASI: intelligence-quotient score (StdS)†	91.0 (89.0, 93.0)	105.8 (104.9, 106.6)	112.7 (111.6, 113.8)	–1.2 (–1.4, –1.1)	–0.1 (–0.1, –0.0)	0.5 (0.4, 0.6)
WASI: matrix reasoning subtest (TS)	44.4 (43.1, 45.7)	52.5 (51.9, 53.1)	56.4 (55.7, 57.2)	–1.1 (–1.3, –0.9)	–0.1 (–0.1, 0.0)	0.4 (0.3, 0.5)
WASI: vocabulary subtest (TS)	44.8 (43.3, 46.3)	54.2 (53.6, 54.9)	58.3 (57.4, 59.1)	–1.1 (–1.3, –0.9)	–0.0 (–0.1, 0.0)	0.4 (0.3, 0.5)
WRAML2: attention/concentration index (StdS)	89.6 (87.2, 91.9)	101.9 (100.9, 103.0)	113.0 (111.6, 114.3)	–1.1 (–1.2, –0.9)	–0.2 (–0.2, –0.1)	0.6 (0.5, 0.7)
CPT-II: # commissions (TS)	55.0 (53.1, 57.0)	53.9 (53.1, 54.8)	44.0 (42.9, 45.1)	0.4 (0.2, 0.5)	0.3 (0.2, 0.4)	–0.6 (–0.7, –0.5)
CPT-II: # omissions (TS)	55.8 (52.9, 58.6)	55.3 (54.1, 56.5)	48.5 (46.9, 50.1)	0.2 (0.1, 0.3)	0.2 (0.1, 0.2)	–0.3 (–0.4, –0.2)
CPT-II: hit RT (TS)	56.0 (54.1, 57.8)	49.8 (49.0, 50.7)	53.8 (52.8, 54.9)	0.4 (0.3, 0.6)	–0.2 (–0.3, –0.1)	0.2 (0.1, 0.3)
CPT-II: hit RT standard error (TS)	56.4 (54.7, 58.2)	50.8 (50.1, 51.6)	47.0 (46.1, 48.0)	0.7 (0.5, 0.8)	0.1 (–0.0, 0.1)	–0.3 (–0.4, –0.2)
CPT-II: variability (TS)	56.2 (54.5, 58.0)	51.6 (50.8, 52.4)	46.3 (45.4, 47.3)	0.6 (0.4, 0.8)	0.1 (0.1, 0.2)	–0.4 (–0.5, –0.3)
CPT-II: detectability (TS)	55.1 (53.4, 56.8)	54.2 (53.5, 55.0)	46.0 (45.0, 47.0)	0.3 (0.2, 0.5)	0.3 (0.2, 0.4)	–0.6 (–0.7, –0.5)
WRAML2: verbal memory index (StdS)	93.2 (91.2, 95.3)	104.6 (103.7, 105.4)	114.0 (112.9, 115.2)	–1.1 (–1.2, –0.9)	–0.2 (–0.2, –0.1)	0.6 (0.5, 0.7)
WRAML2: story memory delay recall subtest (SS)	9.2 (8.8, 9.6)	11.3 (11.1, 11.5)	12.5 (12.3, 12.7)	–1.0 (–1.2, –0.8)	–0.1 (–0.1, 0.0)	0.4 (0.3, 0.5)
WRAML2: verbal learning delay recall subtest (SS)	8.5 (8.0, 8.9)	10.0 (9.8, 10.2)	12.2 (11.9, 12.4)	–0.8 (–0.9, –0.6)	–0.2 (–0.3, –0.1)	0.6 (0.5, 0.7)
Delayed verbal recall composite (StdS)†	94.1 (92.4, 95.9)	103.3 (102.5, 104.0)	111.6 (110.6, 112.6)	–1.0 (–1.2, –0.9)	–0.2 (–0.2, –0.1)	0.6 (0.5, 0.7)
WRAML2: story memory recognition subtest (SS)	9.9 (9.5, 10.3)	11.5 (11.4, 11.7)	12.3 (12.1, 12.6)	–0.7 (–0.9, –0.5)	–0.0 (–0.1, 0.0)	0.3 (0.2, 0.4)
WRAML2: verbal learning recognition subtest (SS)	8.8 (8.4, 9.2)	10.5 (10.3, 10.6)	11.6 (11.4, 11.8)	–0.8 (–1.0, –0.6)	–0.1 (–0.2, –0.0)	0.4 (0.3, 0.5)
Verbal recognition composite (StdS)†	96.7 (95.1, 98.4)	105.0 (104.3, 105.7)	109.9 (108.9, 110.8)	–1.0 (–1.1, –0.8)	–0.1 (–0.2, –0.0)	0.4 (0.3, 0.5)
WRAML2: design memory subtest (SS)	6.2 (5.8, 6.7)	8.5 (8.3, 8.7)	10.5 (10.2, 10.7)	–1.0 (–1.2, –0.8)	–0.1 (–0.2, –0.1)	0.6 (0.5, 0.7)
WRAML2: design recognition subtest (SS)	7.8 (7.3, 8.4)	9.7 (9.5, 10.0)	11.4 (11.1, 11.7)	–0.7 (–0.9, –0.6)	–0.1 (–0.2, –0.1)	0.4 (0.3, 0.5)
D-KEFS trail making test: condition 1 (SS)	7.8 (7.4, 8.3)	10.9 (10.7, 11.1)	11.6 (11.3, 11.8)	–1.1 (–1.3, –1.0)	0.0 (–0.0, 0.1)	0.3 (0.2, 0.4)
D-KEFS trail making test: condition 2 (SS)	8.9 (8.5, 9.3)	11.4 (11.2, 11.6)	12.3 (12.0, 12.5)	–1.1 (–1.3, –0.9)	–0.0 (–0.1, 0.1)	0.4 (0.3, 0.4)
D-KEFS trail making test: condition 3 (SS)	7.0 (6.6, 7.5)	11.3 (11.1, 11.5)	12.4 (12.2, 12.7)	–1.5 (–1.6, –1.3)	0.0 (–0.0, 0.1)	0.4 (0.3, 0.5)
D-KEFS trail making test: condition 4 (SS)	5.9 (5.5, 6.4)	10.1 (9.9, 10.3)	11.7 (11.4, 11.9)	–1.5 (–1.7, –1.3)	–0.0 (–0.1, 0.0)	0.5 (0.4, 0.6)
D-KEFS trail making Test: condition 5 (SS)	10.0 (9.7, 10.3)	11.5 (11.4, 11.6)	12.1 (11.9, 12.3)	–0.9 (–1.1, –0.7)	–0.0 (–0.1, 0.1)	0.3 (0.2, 0.4)
D-KEFS tower test: total achievement score (SS)	8.2 (7.8, 8.6)	10.2 (10.1, 10.4)	10.4 (10.2, 10.6)	–0.8 (–1.0, –0.6)	0.1 (–0.0, 0.2)	0.1 (0.0, 0.2)
Wisconsin Card Sort: perseverative responses (TS)	46.1 (43.8, 48.4)	53.7 (52.7, 54.7)	63.1 (61.8, 64.4)	–0.7 (–0.9, –0.6)	–0.2 (–0.3, –0.1)	0.5 (0.4, 0.6)
Wisconsin Card Sort: perseverative errors (TS)	45.6 (43.5, 47.7)	53.1 (52.2, 54.1)	62.0 (60.8, 63.2)	–0.8 (–0.9, –0.6)	–0.2 (–0.3, –0.1)	0.6 (0.5, 0.7)
CTOPP: rapid naming composite score (StdS)	88.1 (85.5, 90.7)	95.7 (94.6, 96.8)	100.6 (99.1, 102.0)	–0.6 (–0.8, –0.4)	–0.1 (–0.1, 0.0)	0.3 (0.2, 0.4)
D-KEFS verbal fluency: category fluency (SS)	9.6 (9.1, 10.2)	11.6 (11.3, 11.8)	12.7 (12.4, 13.0)	–0.7 (–0.9, –0.5)	–0.1 (–0.1, 0.0)	0.3 (0.2, 0.4)
Beery: motor coordination (StdS)	72.2 (69.9, 74.4)	86.2 (85.2, 87.2)	93.5 (92.2, 94.7)	–1.1 (–1.3, –0.9)	–0.1 (–0.1, 0.0)	0.5 (0.4, 0.6)
Grooved Pegboard dominant hand (StdS)	72.4 (69.1, 75.6)	99.0 (97.6, 100.4)	105.9 (104.1, 107.7)	–1.3 (–1.5, –1.2)	0.0 (–0.1, 0.1)	0.4 (0.3, 0.5)
Grooved Pegboard other hand (StdS)	65.5 (61.6, 69.4)	96.3 (94.6, 98.0)	104.9 (102.7, 107.1)	–1.3 (–1.5, –1.1)	0.0 (–0.1, 0.1)	0.4 (0.3, 0.5)
Fine motor study composite (StdS)†	72.3 (70.1, 74.4)	92.6 (91.7, 93.5)	99.7 (98.5, 100.9)	–1.5 (–1.7, –1.3)	–0.0 (–0.1, 0.0)	0.5 (0.4, 0.6)
Beery: visual-motor integration (StdS)	76.7 (74.5, 78.8)	87.4 (86.5, 88.4)	95.0 (93.8, 96.1)	–1.0 (–1.2, –0.8)	–0.1 (–0.2, –0.0)	0.5 (0.4, 0.6)
Beery: visual perception (StdS)	91.0 (89.2, 92.9)	98.7 (97.9, 99.5)	102.5 (101.4, 103.5)	–0.8 (–1.0, –0.6)	–0.0 (–0.1, 0.0)	0.3 (0.2, 0.4)

Test scores for each cluster were also expressed as differences from the normative population means for each test (Supplementary file S5). For Cluster A, the means of several scores depending on motor skills were more than 1 sd lower than the population norms. For Cluster C, the means of scores on the Wisconsin Card Sort (a measure of executive function) were greater than 1 sd higher than the population norms. All other scores were within 1 sd of the population norms.

Subjects in Cluster A were more likely to have younger mothers, parents with lower educational attainment, to be male, have a lower gestational age, have a lower 5 min Apgar, lower birthweight, and lower socioeconomic status (Supplementary file S6). Parent reports of behavioural and learning problems differed significantly according to cluster membership (Table 4), with parents reporting more problems in subjects belonging to Cluster A and fewer problems in subjects belonging to Cluster C.

Table 4.

Values of parental reports according to cluster membership.*Analysis of variance F-test. ^†χ². ADHD, attention deficit hyperactivity disorder; BRIEF, Behavior Rating Inventory of Executive Function; CBCL, Child Behavior Checklist; CLDQ, Colorado Learning Difficulties Questionnaire; sd, standard deviation. ZS, Z-score [mean=0; sd=1]; TS, T-score (mean=50, sd=10). For each characteristic, higher mean scores indicate more problems. For the BRIEF and CBCL, scores >60 are considered to be clinically significant.

Characteristic	Cluster A (n=106)	Cluster B (n=557)	Cluster C (n=334)	P-value
CLDQ: math scale (ZS)				<0.001*
Missing	8	71	24
Mean (sd)	1.44 (1.33)	0.11 (1.12)	–0.36 (0.79)
Range	–0.86, 3.59	–0.86, 3.59	–0.86, 2.92
CLDQ: reading scale (ZS)				<0.001*
Missing	8	71	23
Mean (sd)	0.98 (1.30)	–0.14 (0.89)	–0.46 (0.62)
Range	–0.81, 3.63	–0.81, 3.08	–0.81, 2.71
BRIEF: global executive composite				<0.001*
Missing	10	105	42
Mean TS (sd)	56.9 (12.8)	48.3 (10.0)	45.1 (9.0)
Range	37, 92	31, 87	31, 71
Number (%) with score >60	31 (31.0%)	59 (11.3%)	25 (7.6%)	<0.001†
CBCL category: internalising problems				<0.001*
Missing	14	110	44
Mean TS (sd)	52.9 (12.9)	48.3 (10.4)	46.9 (9.4)
Range	33, 87	33, 80	33, 72
Number (%) with score >60	25 (26.0%)	61 (11.8%)	28 (8.6%)	<0.001†
CBCL category: externalising problems				<0.001*
Missing	14	110	44
Mean TS (sd)	50.2 (11.4)	45.1 (10.1)	42.5 (8.6)
Range	33, 86	33, 78	31, 69
Number (%) with score >60	18 (18.8%)	33 (6.4%)	12 (3.7%)	<0.001†
CBCL category: total problems				<0.001*
Missing	14	110	45
Mean TS (sd)	53.4 (12.3)	45.6 (11.3)	42.1 (10.3)
Range	26, 86	24, 77	24, 68
Number (%) with score >60	23 (24.0%)	45 (8.7%)	13 (4.0%)	<0.001†
CBCL category: ADHD problems				<0.001*
Missing	14	109	46
Mean TS (sd)	57.4 (8.1)	53.4 (5.9)	51.6 (3.8)
Range	50, 78	50, 77	50, 75
N (%) with score >60	26 (27.1)	52 (10.0)	11 (3.4)	<0.001†

Of the multiply exposed subjects, 23.3% belonged to Cluster A. The weighted (using IPTW) proportions of unexposed and singly exposed children in Cluster A were 10.1% and 9.8%, respectively (Table 5). In weighted multinomial logistic regression, exposure status was significantly associated with cluster membership (overall P=0.002). The odds of multiply exposed children belonging to Cluster A was 2.83 (95% CI: 1.49–5.35; P=0.001) higher than unexposed children compared with belonging to Cluster B (Table 5). There were no other significant associations between exposure status and cluster membership. There was no evidence for interactions between exposure status and any baseline characteristics on cluster membership, as no interaction term was statistically significant (data not shown). In a post hoc sensitivity analysis, removal of the 18 children who underwent cardiovascular or intracranial procedures did not affect the association between exposure status and cluster membership (data not shown).

Table 5.

Anaesthesia exposure and cluster membership. *Inverse probability of treatment weighting was used to account for imbalances in covariates across exposure categories. The values in the table represent the proportion of patients in each cluster by exposure status after weighting. ^†Overall P-value for the exposures variable. P<0.05 indicates that exposure status was significantly associated with cluster membership. ^‡P=0.001, multinomial linear regression.

	Weighted proportion within each exposure status* (%)			Odds ratios (95% confidence interval)		P-value†
	Cluster A	Cluster B	Cluster C	Cluster A vs Cluster B	Cluster C vs Cluster B
Exposures						0.002
None	10.1	58.2	31.3	1.00 (reference)	1.00 (reference)
Single	9.8	57.6	32.7	1.02 (0.50, 2.09)	1.06 (0.69, 1.63)
Multiple	23.3	48.1	28.6	2.83 (1.49, 5.35)‡	1.09 (0.72, 1.67)

Discussion

The findings of this secondary analysis of data from the MASK study support the hypothesis that exposure to multiple, but not single, procedures requiring general anaesthesia is associated with a specific pattern of differences in neuropsychological tests in some children that may have consequences for subsequent learning and behaviour.

Factor analysis

Factor analysis is useful to analyse a large number of variables in terms of any common underlying unobserved relationships (denoted as factors). It is commonly utilised in psychological research as a technique to address the challenges associated with making multiple individual comparisons of potentially correlated variables, as was done in the original analysis of the MASK data set.²⁴ EFA is used to identify these factors based on shared variability of measurements (in this case, neuropsychological tests) without a priori assumptions regarding any such relationships. Once hypothesised relationships were identified in a random sample of half the subjects, CFA was used to test this hypothesised model in the remaining subjects. The five factors identified were generally interpretable in terms of the putative domains tested by the individual neuropsychological tests. Three of the factors (Factors 3–5) included scores from a single test. One factor (Factor 2) included measures of verbal concentration, learning, and memory. The remaining factor (Factor 1) primarily included tests dependent on processing speed, motor coordination, and visual-motor integration.

An analysis of the association between anaesthesia exposure status and factor scores supported the hypothesis generated by the prior analysis of individual tests.²⁴ There was no evidence that exposure is associated with differences in factors related to verbal and memory skill, attention, impulsivity, and many measures of executive function. This is consistent with the results of a prior analysis, which also found no evidence of effects when each individual test evaluating these domains was considered.²⁴ In those multiply exposed, there was a modest (∼1/3 sd) decrease in the score of Factor 1, which was comprised of several scores that were also significantly associated with multiple exposures in the prior analysis [including Delis–Kaplan Executive Function System Conditions 1–3 and 5, Comprehensive Test of Phonological Processing, Beery motor coordination, and Grooved Pegboard], and others that suggested associations, but were not statistically significant (WASI matrix reasoning subtest, WRAML2 attention/concentration index, and Beery visual-motor integration and visual perception). As noted previously, deficits in these domains are often observed in children with attention deficit hyperactivity disorder (ADHD) or reading difficulties,37, 38, 39 which are more common in multiply exposed children in this population.8, 10, 22

Cluster analysis

The prior analysis examined mean effects over all children tested²⁴; this approach may not be sufficiently sensitive to detect significant effects in other domains if only some children are affected.²⁵ One method to evaluate this possibility is cluster analysis, useful to define groups of subjects similar in some way—in this case, pattern of performance on neuropsychological tests. The resulting clusters grouped subjects largely according to overall performance on most of the tests, identifying subjects with relatively higher and lower performances. However, we caution against interpreting membership in these clusters as indicating ‘abnormal’ or ‘superior’ performances for individual subjects. First, within each cluster, there was variability in scores, such that individual scores for individual subjects could overlap between clusters. Second, many neuropsychological scores were within 1 sd of the normative sample, raising the possibility that such findings may have little effect upon functioning. Third, in terms of interpreting function, differences from the overall population mean, rather than the mean of all subjects sampled, may be of greater importance. The study population was selectively sampled according to the propensity for receiving anaesthesia, and thus, does not necessarily reflect the characteristics of the overall population. For the 31 test scores, the means of all but three scores for children in the lower-performing Cluster A were less than the population mean. For eight of these test scores, all of which depend on motor coordination, performance was quite poor for children in Cluster A, approaching or exceeding 2 sd below the population mean.

Cluster membership was consistent with known determinants of traits, such as intelligence, including parental education, socioeconomic status, and birthweight,40, 41 supporting the validity of the clustering procedure. Cluster membership was also associated with parental reports of behavioural and learning problems, both as continuous variables and as the proportion of subjects above a threshold value thought to indicate clinically significant abnormalities; approximately one-fourth of subjects in the lowest-performing cluster were rated as having clinically significant behavioural problems. This finding is consistent with known associations between deficits in measures, such as motor skills and processing speed, and behavioural and learning difficulties,38, 42, 43 and suggests that the observed differences in neuropsychological tests have consequences for behaviour and learning. This finding also suggests that learning and behavioural problems tend to occur in the same children, consistent with prior observations of the co-occurrence of learning disability and ADHD, especially in children with multiple exposures to anaesthesia.10, 22, 44, 45

A greater proportion of children who were multiply exposed were in the lower-performing cluster. Given that subjects in this cluster had scores less than the population means for almost every test, it is not possible to exclude the possibility that multiple exposures may be associated in some children with deficits not only in processing speed and motor coordination, but other domains as well, deficits insufficient to produce significance when analysed across the entire study population. However, it is important to note that most of the multiply exposed children were in the two higher-performing clusters. In these two clusters, scores on almost all tests (excepting the Berry motor coordination and visual-motor integration for both Clusters B and C, and the fine motor study composite and Grooved Pegboard other hand for Cluster B) and parent reports were at or above population norms. This finding suggests that the majority of even multiply exposed children do not exhibit a pattern of clinically significant deficits. The analysis did not suggest any moderator of associations, so that it remains unclear why some children and not others show this pattern.

Study limitations have been extensively discussed in previous work, primarily the potential for unmeasured confounders to affect outcomes (e.g. confounding by indication).22, 23, 24, 46 Fundamentally, children who need procedures differ from those who do not, and such differences may not be fully accounted for in analysis. The finding of a specific pattern of differences may argue against confounding by indication, as it is not clear what common underlying condition could produce a specific pattern. Nonetheless, these analyses cannot directly demonstrate causality. Other potential limitations include (i) selection bias, because not all invited agreed to participate²⁴; (ii) characteristics of Olmsted County residents, which differ from the US population as a whole; (iii) potential for exposures after 3 yr of age to affect results; and (iv) potential for factors associated with the procedure itself, rather than anaesthesia, to produce causal effects. Finally, factor and cluster analyses themselves have their own limitations, including sensitivity of results to assumptions and choices made in their conduct,29, 30, 31, 33, 35 such that factor and cluster structures are not unique.

Conclusions

Multiple, but not single, exposures to procedures requiring general anaesthesia before age 3 yr are associated with a specific pattern of deficits in some neuropsychological tests. Factors predicting which children develop the most pronounced deficits remain unknown.

Authors' contributions

Study design: MJZ, DRS, RTW, JS, SKK, RPF, DOW.

Obtaining funding: RPF, DOW.

Data collection: MJZ, DRS, DOW, RPF.

Data analysis: MJZ, RDF, ML, PJS, DRS, ACH, DOW.

Data interpretation: MJZ, RDF, ML, PJS, ACK, RTW, JS, SKK, RPF, DOW.

Drafting paper: DOW.

Revising paper: MJZ, RDF, ML, PJS, ACK, DRS, ACH, SLB, RTW, JS, SKK, RPF

Final version approval: all authors.

Agreement to accountability: all authors.

Handling editor: H.C. Hemmings Jr

Editorial decision: 18 January 2019

Declaration of interest

The authors declare no financial and personal relationships with other people or organisations that could inappropriately influence (bias) their work.

Funding

Eunice Kennedy Shriver National Institute of Child Health and Human Development of the National Institutes of Health (R01 HD071907); National Institute on Aging of the National Institutes of Health (R01 AG034676).

Appendix A. Supplementary data

The following are the Supplementary data to this article: