Annual Research Review: The state of autism intervention science: progress, target psychological and biological mechanisms and future prospects

Trial methods

An immediately striking feature of the contemporaneous systematic review by French and Kennedy (2018) is how few of the studies they identified met basic Cochrane quality criteria for absence of risk of bias (in just six trials out of the 48 that met their inclusion criteria, even though most had been conducted in the last few years). Another is how chaotic the autism intervention field is internationally, with those 48 trials testing no less than 34 different models of intervention, usually in underpowered small N trials from single sites.

While there has been substantive progress in recent years, it remains striking and paradoxical that the intervention science in such an important developmental disorder continues to be so fragmented and methodologically fragile. This cannot just be put down to lack of funding. The majority of interventions tested have been behavioural or psychosocial, areas in which there has been debate (historically but also currently) as to the appropriateness or viability of conducting rigorous trials. Any doubts as to the value and importance of Randomised Controlled Trials (RCT) methodology in this area however should have been laid to rest. It turns out that they are not only feasible but can have great advantages in clarifying and adjusting for the many potential confounds involved in properly testing complex interventions within an equally complex phenotype (Dunn et al., 2015). It has been a genuine achievement for the field (in parallel certainly with other areas of psychosocial treatment research) to have demonstrated the value of RCT methods and, increasingly, the implementation of CONSORT reporting standards; but the French and Kennedy (2018) review shows how patchy this has been in practice. Enormous sums of public and private money are spent in health care systems across the world on interventions that have no rigorous evidence base. For instance, probably the most ubiquitous interventions globally (under the umbrella of Early Intensive Behavioural Intervention (EIBI) or ABA) continue to be actively promoted on the basis of single case designs alone, internationally considered to be low levels of evidence in healthcare; a Cochrane review of EIBI (Reichow, Barton, Boyd, & Hume, 2012) found only one historical, small and equivocal RCT and the situation has not changed subsequently. Consequently, such interventions and others do not feature at all in the French and Kennedy (2018) review or in relevant guidance such as UK NICE (National Institute for Health & Care Excellence, 2013). This is a truly troubling situation given the high profile and importance of autism in public health, and its economic and social costs. On just one metric, the life span economic cost of autism per individual has been estimated as up to $2.4 million in the United States and £1.5 million in the United Kingdom (Buescher, Cidav, Knapp, & Mandell, 2014; Knapp, Romeo, & Beecham, 2009), higher than asthma or diabetes. The reader might imagine if international clinical practice in these latter conditions was underpinned in this way.

In a further paradox, the autism field actually has a significant advantage over some other areas of psychosocial intervention research in having developed relatively valid, predictive and blind-codable measures of social interaction and autism symptom outcomes; something that has often been elusive in other areas. Lack of assessment blinding in self- or informant- reporting can seriously inflate treatment effect estimates, especially in trials without participant blinding to treatment, inevitably the norm in most psychosocial treatments (Sonuga-Barke et al., 2013). In addition to incorporation of common blinded outcomes, a further key next step in methodological development will be widespread implementation of pre-recruitment trial registration and pre-specification of analysis plans and primary outcomes, avoiding the post hoc selective reporting of multiple measured outcomes that has often been the norm. At a stroke this discipline, although arduous initially, would transform the validity and salience of the trial reporting.

A third key step will be the explicit investigation of mechanism, which will allow an iterative evolution towards more refined treatments (Green, 2015; Kraemer, Wilson, Fairburn, & Agras, 2002). The first part of this review consequently builds on French and Kennedy (2018) by focusing on the (explicit or implicit) mechanisms and treatment targets within the variety of treatments tested to date – particularly in those studies reviewed by French and Kennedy (2018) as having relatively low risk of bias. We aim through this to provide a unifying view of the current research along with an agenda for moving forward with new and refined mechanism targets. We then point the way forward to some of these new potential targets in brain science in the second part of the review.

Co-construction of outcome targets

A further achievement for autism intervention has been an emerging dialogue with user groups, advocates and families towards coconstructing relevant outcome variables and outcome targets after intervention. Autism has benefitted from active debates and dialogue between these groups, which should be creative for future progress. At the extreme some advocates object to any idea of intervention at all as impacting negatively on ‘autism identity’, wanting instead a focus on acceptance, social rights and associated mental health. These latter objectives are of undeniable importance, but the review below addresses how interventions, theoretically targeted at core early developmental processes, can impact positively on key autism features and adaptation, even in this highly heritable neurodevelopmental disorder. This empirical evidence, alongside early diagnosis, forms the ethical case for intervention; but the outcomes selected do need to be of shared relevance. This is a theme we return to at the end of the review.

Parent-child dyadic interaction as a foundation for social development

The rationale for targeting early parent-child social interaction (PCI) in autism as a social impairment disorder comes from decades of child development research suggesting the foundational value of positive early dyadic social interaction for later child social communication and adaptation (Landry, Smith, Swank, & Miller-Loncar, 2000; Page, Wilhelm, Gamble, & Card, 2010; Tomasello, 2008). Alongside this, autism developmental science has identified variations from such normative patterns in young children at risk of or diagnosed with the condition. Thus increased parent directiveness (but not reduced responsiveness) is found in the latter part of the first year in infant siblings at risk of autism (Harker, Ibanez, Nguyen, Messinger, & Stone, 2016; Wan et al., 2013) along with reduced child attentiveness to parent, affective signalling and coordinated dyadic communication (Parlade & Iverson, 2015; Wan et al., 2013). Similar patterns are seen in established autism as well as in other developmental disabilities (Blacher, Baker, & Kaladjian, 2013; Doussard-Roosevelt, Joe, Bazhenova, & Porges, 2003). We cannot of course assume that such patterns of PCI are responsible for autism outcomes; it could be in theory that these different PCI trajectories are incidental or even adaptive (Mottron, 2017). Nevertheless, a transactional account (Sameroff, 2009) of the early reciprocal interplay between intrinsic developmental atypicalities in autism and the early social environment would suggest that evoked perturbations in interaction can indeed serve to amplify social impairments and symptom impact over time; on this basis, reversing these deficits and optimising (‘normalising’) social interaction could in theory improve social functioning or reduce symptoms in autistic children. The PCI target has the added theoretical advantage of being ‘naturalistic’ (i.e. working with the everyday developmental/family system of the child) and potentially more easily generalisable than clinic-based interventions directly with the child. It is however also based on the assumption that intervening in the developmental system in this way will be ‘powerful’ enough significantly to impact autism development, traditionally seen to be intrinsic and relatively environmentally stable. Whether this assumption holds is an empirical question to which we will return.

Interventions designed to impact PCI have often used demonstration and modelling by the therapist working directly with the child in the parent's presence or mixed with didactic education. More recently, there has been a rise in more exclusive work with the parent using direct coaching in real-time during interaction or use of video-feedback techniques. The precise ‘proximal’ intervention targets within PCI (that is, the immediate goal for change with intervention) have varied from parental ‘responsiveness’ and ‘non-directiveness’, parental ‘communicative synchrony’, to mutual ‘joint attention and engagement’ and ‘affect sharing’ (see Figure 1). Each of these targets implies a slightly different theoretical orientation and defined goals.

Behavioural training methods

Behavioural learning methods use focused behaviour modification techniques to influence behaviour change. In traditional behavioural interventions for autism the targets are reduction in unwanted (often repetitive or assumed non-socially functional) behaviours along with increase in identified social behaviours. In more recent studies, the targets for behavioural change are more developmentally informed; for instance, in pivotal response therapy (PRT) the targets are behaviours such as motivation for engagement and interaction considered theoretically ‘pivotal’ for generating wider generalised change. Broader ‘curriculum’ models such as Early Start Denver Model (ESDM) select a range of developmentally relevant targets, from communication to play to motor activity. What distinguishes these models however is that the methods used to achieve therapeutic change towards such targets are still taken from behavioural learning theory, largely using repetition and contingency reinforcement for desired behaviours (Koegel, Singh, Koegel, Hollingsworth, & Bradshaw, 2014). While use of contingencies is part of naturalistic parenting, its use in these models is much more exclusive and intensively focused. Interventions of this kind are typically intensive and therapist-child delivered over several years. The only trial of this type of intervention identified by French and Kennedy (2018) as having acceptable low risk of bias was, however, of a relatively low intensity (one session/week over 12 weeks) group PRT intervention, focused on improving functional communication (Hardan et al., 2015). The primary outcome was the frequency of functional child utterances in the dyad coded during structured lab interaction in which parents were instructed to ‘try getting the child to communicate as much as possible’. Parents were rated as implementing PRT methods successfully; total child utterances increased in the PRT group, with the effect driven by imitative and non-verbally prompted utterances rather than spontaneous utterances; and verbal prompts had no effect. The pattern of this result could, however, be subject to ‘training to the test’; parental contingent reinforcement of child utterances, trained in the therapy, produced local context-specific increase in the target child behaviour. Evidence for generalisation is modest; independent global impression of social communication from same tapes alongside parent interview reported improvements as did parent-reported communication on Vineland, but there was no effect in parent-reported language use, standard language measures or parent-rated autism behaviours. A subsequent similar intervention model combining PRT and ABA reported effects on dyadic joint engagement and shared enjoyment but also no generalised effect (Brian, Smith, Zwaigenbaum, & Bryson, 2017).

Summary - proximal dyadic targets

The evidence outlined above shows that caregiver dyadic behaviours with their child (Figure 1A) are notably responsive to a number of different interventions (just as they can be in non-autism contexts). This applies across the spectrum of social class, ethnicity and education. Replicated effects from robust trials of various interventions in autism show moderate to large effect sizes, particularly to improve parental responsiveness sensitivity and synchrony, theoretically linked in development to positive child social and communication outcomes. In relation to intervention techniques, video feedback probably has the strongest evidence for producing reliable parental change in this kind in both autism (Aldred, Green, & Adams, 2004; Green et al., 2010; Kasari, Siller et al., 2014; Poslawsky et al., 2015; Rahman et al., 2016) and neurotypical groups (Juffer, Bakermans-Kranenburg, & van IJzendoorn, 2008). Techniques primarily using group or individual coaching and modelling techniques with parents do not show such reliable results (Carter et al., 2011; Rogers et al., 2012). Targets involving reciprocal dyadic interaction between parent and child, such as joint engagement and joint attention (Figure 1B) have also proved treatable in replicated studies. There is less consistent evidence that the child's dyadic social interaction in turn can be improved (Figure 1C). Many studies demonstrating increased parental responsiveness have not found that this transmitted well into child effects; others have shown change in both parental responsiveness and child-initiated joint engagement joint attention (Shire et al., 2017). However, focused video-feedback for parents in the PACT and iBASIS trials and a mixed coaching, modelling and video work in JASPER are all associated with reciprocal effects on spontaneous child communication in the treatment dyad.

Convergent evidence thus demonstrates that immediate PCI targets are amenable to intervention. However, such targets are highly proximal to the intervention effort, and in many studies these outcomes are measured within or in very similar contexts to the treatment setting, potentially confounding independence and making possible a simple ‘training to the test’ result. It is moreover a feature of autism development that the generalisation of acquired skills across contexts is particularly difficult: it cannot be assumed that short-term changes in dyadic interactions of this kind will translate well into more generalised social gains for the child. In this sense, the majority of autism intervention science to date (and the vast majority of published papers) has not really moved beyond what are in effect the equivalent of pre-clinical or early phase studies looking at potential developmental mechanisms, on the (implicit or explicit) assumption that these will go on to effect more generalised autism development and functioning. Even if well done, we can hardly infer from such evidence that a treatment is ‘effective for autism’: and the (frequent) claims to this effect are inflated and unjustified.

Concurrent child outcomes beyond the dyad

By contrast, measuring child concurrent outcomes beyond the immediate dyadic context (Figure 1D) represents a first stage of true generalisation for interventions whose proximal focus is the dyad. It also represents the immediate outcome for therapist-child interventions without parental involvement. An example of such concurrent child social outcome is the behavioural symptoms that define autism disorder itself.

Downstream developmental targets – symptoms and adaptation

There has been a relative lack of developmental follow-on studies after early intervention (Figure 1E), although the situation is now beginning to change. The results are somewhat encouraging but have been mixed. On relatively short-term follow-up, Poslawsky et al. (2015) found sustained treatment effect on child initiation of joint attention at 3 months after the video feedback VIPP-AUTI intervention, but not on other measures. Kasari, Paparella, Freeman, & Jahromi, 2008 showed effect on objective language outcomes from the JASPER intervention using the Reynell Developmental Language Scales (RDLS) at 12-month follow-up but this was not replicated in a later study (Kaale, Fagerland, Martinsen, & Smith, 2014). Chang et al., 2016 measured uncontrolled follow-up of the intervention group only and reported mixed findings. Kasari, Lawton et al. showed maintenance of improvement in their CMM group at 3-month follow-up, but no between-group effect. Other controlled studies that showed between-group endpoint effects on either parent or child dyadic social behaviours do not see them at short or medium-term follow-up (Baranek et al., 2015; Kasari, Siller et al., 2014).

Four RCTs have however now reported longer-term (>1 year) autism symptom outcomes after preschool intervention. Estes et al. (2015) measured ADOS total score at a mean age of 6 years, 2 years after the end of a 2-year intensive developmental behavioural intervention (ESDM). While there had been no effect on symptoms at trial endpoint, the 21 of 24 children in the intervention group followed up showed greater total and RRB symptom reduction over the follow-up period compared to 18 of 24 followed from the control group. Causal treatment effect inference from these findings is limited however by the lack of intention to treat (ITT) analysis or adjustment for baseline pretreatment variables. Kaale et al. (2014) reported no effect on non-blinded parent- or teacher-rated autism symptoms (SCQ) 12 months after a 12-week teacher-mediated JASPER intervention (n = 33) compared to controls (n = 27). In contrast, follow-up of the original PACT trial, 6 years after treatment endpoint, assessed 121 of the original 152 trial participants with core autism at mean age 10.5 years. ITT analysis showed reduction in autism symptom severity (ADOS CSS) at both treatment endpoint (ES 0.64, 95% CI 0.07, 1.20) and follow-up (ES 0.70, 95% CI −0.05, 1.47), resulting in a moderate averaged treatment effect on symptoms over the total period (ES 0.55, 95% CI 0.14, 0.91). Non-blind parent-rated autism symptoms on SCQ (ES 0.40, 95% CI 0.05, 0.77) and repetitive behaviours on RBQ (ES 0.87, 95% CI 0.47, 1.35) also showed comparable improvement at follow up (Pickles et al., 2016). A similar video-feedback social communication intervention for infants at-risk of autism, implemented between 9 and 14 months of age and followed at one and then 2 years from end of treatment (Green et al., 2015), showed interestingly parallel ITT results, with a significant effect of intervention over the treatment and follow-up period for autism prodromal symptoms (ES = 0.32; 95% CI 0.04, 0.60; p = .026), parental dyadic non-directiveness/synchrony (ES = 0.33; CI 0.04, 0.63; p = .013), and child attentiveness/communication initiation (ES = 0.36; 95% CI 0.04, 0.68; p = .015).

Mediation testing in psychosocial intervention

Only two studies in the extant autism literature have mounted a substantive test of mechanism through a mediation test within an RCT. Gulsrud, Hellemann, Shire, & Kasari (2016) used videotape analysis of the content of therapeutic sessions in the JASPER intervention on key putative active ingredients. There was a possible methods confound in that mediator and primary outcome (joint engagement) were both coded (albeit independently) at endpoint from the same video-material with no midpoint assessment, allowing the possibility of ‘reverse causation’ in the inference of effect. Nevertheless, the analysis did identify a particular aspect of the therapy (parental ‘mirrored pacing’) as mediating the joint engagement outcome; which is an important result with generalisable value and points forward to potential further refinements of the therapy procedure. Within the PACT trial, Pickles et al. (2015), identified a latent midpoint mediator variable using baseline, midpoint and endpoint repeated measures, a method of analysis that also reduces the measurement error intrinsic to observational video coding. Treatment effect on parent synchrony strongly mediated (90%) the change found in child communication initiations with parent, supporting the treatment theory. Both these mediation analyses are thus convergent in identifying similar specific parental interaction behaviours (‘mirrored pacing’ and ‘communicative synchrony’) that improve child engagement and communication in the dyadic context. They stand as the clearest mechanism findings to date in autism intervention science and provide important and generalisable information about what are the active components of the therapies in facilitating interaction and child outcomes. Furthermore, the Pickles et al. analysis also showed that it was the change in child dyadic communication initiations with parent (rather than the parental synchrony) that mediated the subsequent change in child autism symptoms with a researcher – thereby demonstrating a causal chain of generalisation of child-acquired skill from the PCI context into a researcher-child interaction. More mechanism tests of this kind will be important going forward.

Summary of psychosocial mechanism targets

Autism particularly impacts social functioning in development. We have seen above that a transactional account of social development and reciprocity in autism, recruiting ideas from social-developmental research in neurotypical populations, has led to a rationale for interventions targeting key aspects of social development in autism. The evidence to date shows that a number of such theoretically based and targeted interventions can indeed reliably improve various aspects of the local dyadic social communication of children with autism, particularly with their caregivers (Figure 1A–C, Table 1). There is less work on whether such dyadic change can produce more generalised impact on concurrent within-child outcomes (Figure 1D): there have been positive findings, but the overall evidence is mixed. There is even less work on the crucial issue of long-term developmental outcomes (Figure 1E), but some emerging evidence, especially from the PACT trial, that this is possible over the medium to long term. With a developmental disorder like autism, interventions must surely aspire to such long-term effects in order to be convincing as disorder-specific treatments.

Given the convergent evidence that short-term context-specific proximal change is possible using a variety of intervention methods, further replication of this established result on its own is redundant moving forward (although they are the easiest and cheapest designs to carry out!). It is true that mechanism findings from one trial (Pickles et al., 2015) show a causal link between treatment change on child dyadic interaction with caregiver and generalised symptom change across context, and this is encouraging that these proximal changes may indeed be meaningful in relation to symptoms and generalised adaptation. But such a finding needs replication, and the field now needs larger and more ambitious psychosocial trials that address mechanism, generalisation and the possibilities of long-term developmental and adaptive change, as well as adaptive interventions tailored to patient characteristics and individual response (for instance using sequential multiple assignment randomised ‘SMART’ trials). Accumulation of such trial evidence will refine approaches and further answer an enduring question; what are the limits to clinically relevant effectiveness of ‘environmental’ interventions in a condition that is highly heritable and developmentally persistent?

Neurotransmitters/neuromodulatory system targets

Signalling pathways targets downstream of neurotransmitter receptors

Growth & Neurotrophic factor targets

Deletion/mutation of the SHANK3 gene in Phelan-McDermid syndrome (PMS; also known as 22q13 deletion syndrome) results in reduced expression of scaffolding proteins in the postsynaptic density of excitatory synapses, impairing glutaminergic transmission and synaptic plasticity (Moessner et al., 2007). PMS is associated with 0.5%–2.0% of all ASD cases. Similarly, loss of function mutations of the X-linked gene MECP2 in Rett syndrome affects the structure and function of synapses at the microcircuit level critical for synaptic transmission and plasticity (Guy, Hendrich, Holmes, Martin, & Bird, 2001). Recent studies suggest that Insulin-like growth factor-1 (IGF-1) can have a beneficial effect on synaptic development by promoting neuronal cell survival, synaptic maturation and synaptic plasticity (Bozdagi, Tavassoli, & Buxbaum, 2013). A pilot placebo-controlled double blind RCT using Insulin-like growth factor in nine children with PMS was associated with significant improvements in social impairments and repetitive behaviours (Kolevzon et al., 2014). In Rett syndrome, an open label phase 1 study of IGF-1 in n = 12 girls with MECP2 showed good safety, tolerability and improvements in behavioural abnormalities (Khwaja et al., 2014). Studies of IGF-1 in FXS are also underway.

Other drugs such as Minocycline (a tetracycline derivative approved for treatment of acne) have been shown to have neuroprotective effects and upregulate neurotrophic and growth factors in the brain. In FXS animal models, Minocycline has been shown to have an inhibitory effect on matrix metalloproteinase nine activity which is fundamental in modulating hippocampal plasticity. A double-blind placebo controlled cross-over trial of Minocycline in 55 FXS children aged 3.5–16 years over 12 weeks found significant treatment effect on clinician-rated Clinical Global Impression Scale (CGI) but not on any behavioural measures (Leigh et al., 2013). Furthermore, minocycline use was associated with alteration of event-related potentials in an auditory oddball paradigm in 12 subjects taken from the same trial (Schneider et al., 2013), possibly suggesting reduced hypersensitivity to auditory stimulation.

Electrophysiological targets

At the more general level of neural system functioning and long-range integration and connectivity, EEG and habituation responses to social stimuli have been investigated as targets for ASD intervention. Linked to underlying social processing and social attention, these can be some of the earliest apparent atypicalities in infants at risk who go on to develop ASD (Johnson et al., 2015; Szatmari et al., 2016). As an example of EEG as a target of this kind, Jones, Dawson, Kelly, Estes, & Jane Webb, 2017 reported on the EEG outcomes within an RCT of parent-delivered psychosocial ‘promoting first relationships’ intervention designed to facilitate parent infant interaction and delivered between 9–11 months of infant age. Both at 12-month and 18-month follow-ups they reported increase in social habituation to faces versus objects and greater increase in frontal EEG theta power (Jones et al., 2017). The study was relatively small with valid data available from 15 or less in the PFR group and 10 or less in the control group over the range of assessments, but the findings provide proof of principle and suggestive results that neurophysiological measures can be impacted by relatively low intensity social communication focused intervention. The results echo a previous study (Dawson et al., 2012), which studied EEG in 60% of the trial cases at endpoint only after two years intensive ESDM therapy. Although the findings showed no intervention effect across a number of parameters, there was reported increase EEG-theta to social stimuli with faces, which correlated with a parent-reported social communication rating (although not with autism symptoms, IQ, language or adaptive behaviour).

Summary of neural mechanism targets

Studies of syndromic autism have provided a way of identifying causal links from genes to specific molecular and biological pathways and alterations in behavioural outcomes. An underlying hypothesis is that there may be a finite number of common molecular pathways in autism that could act as intervention targets and that treatments thus identified in the context of syndromic ASD could then be applicable to broader idiopathic ASD; thus raising the possibility of a precision medicine approach to autism based on the underlying pathophysiology (Loth, Murphy, & Spooren, 2016). The work reviewed above reflects the active international effort now in this area, which has already produced some clarification of neural system pathways and encouraging therapeutic leads. These include for instance mTOR inhibitors, statins, oxytocin and pleiotropic growth factors. Despite this, no single treatment has so far been approved for a Phase III trial and there are clearly conceptual as well as practical challenges. Animal experiments cannot fully recapitulate the complexity of higher order cognitive processes in humans. The interventions target a theoretical underlying pathway towards behavioural outcomes, but it is notable that most studies to date have not included detailed measures of this hypothesised pathway and mechanism (Stivaros et al., 2018 is one pilot example that did), and this will be critical in the future. Practical challenges include patient selection, age of treatment onset, dose and duration of treatment, dose-limiting side-effects and lack of CNS biomarkers (Berry-Kravis et al., 2016). Most human intervention work has been in adolescents or adults and it may be that targeting the developing brain in much younger samples may be crucial. Outcome measurement in these studies – commonly the parent-rated Aberrant Behaviour Checklist (ABC; Aman, 1994), or the clinician-rated Global Improvement Scale (CGI; Leucht & Engel, 2006), Table 2 – have also been very different to that in psychosocial research, unfortunately limiting comparability. This is important because to be convincing, treatment studies in this area will need to aspire to the same outcome standards as discussed above for psychosocial research. The need for a common outcomes framework is further discussed below.