Identifying cerebral palsy phenotypes objectively

Many of the definitions of cerebral palsy (CP) acknowledge the heterogeneity and variability of manifestations (e.g. movements, postures, function), and the severity of limitations. This heterogeneity has posed a problem for those trying to identify what contributes to CP occurrence and those who want to assess response to interventions. In 2000, the Surveillance of CP in Europe (SCPE) investigators specifically created diagnostic and classification guidelines that would provide uniformity across populations, clinicians, and investigators. They achieved that to a considerable extent, but variation in assignment of CP subtype remained a problem. Recent advances in molecular genetics have allowed the identification of treatable metabolic disorders associated with CP (e.g. glutaric aciduria type 1 arginanse deficiency). One would expect that microarray and next-generation sequencing studies of CP would pay particular attention to phenotypic definitions of the forms of CP linked with specific genomic and metabolic disorders. However, only 32% of 57 genetic studies of CP defined it on the basis of SCPE international guidelines. This failure to use uniform guidelines limits our confidence in these genetic studies. This is especially unfortunate in light of the increasing use of the term ‘CP mimic’ for gene variations associated with CP syndromes. The report of the potential inadequacies of the genetic studies identified the problem, but did not provide any solution. We now need to begin the discussion as to how physicians who treat CP can collaborate more closely with their geneticist colleagues in the future. Obviously, the first suggestion is to read the study conducted by Pham et al. and the SCPE guidelines published in 2000. Following SCPE guidelines does result in standardization. But standardization does very little to increase confidence in reports of the genetics of CP unless observer variability is also minimized. Unfortunately, subjective spasticity measurement tools do not have sufficiently acceptable psychometric properties (including validity and reliability), nor do assessments of dystonia. The importance of minimizing intraobserver and interobserver variability is emphasized by the results of a recent study of assessments of the presence and severity of spasticity, ataxia, or dyskinesias/dystonia identified in videos of 15 young males with CP. Three pediatric neurologists, three rehabilitation physicians, and three movement disorder specialists viewed the same videos. Nevertheless, these clinicians showed unacceptably high levels of interand intraobserver agreement (even among their colleagues of the same specialty) to instill confidence in efforts to standardize CP diagnosis. This unacceptably high interand intraobserver variability prompted this from the editorial writer: ‘Is there no substitute to actually being there in the examination room?’ These disappointing findings with subjective assessments lead to a reliance on objective assessments. Unfortunately, assessments of motor function in people with dyskinetic CP using 3D motion capture and surface electromyography still do not yet provide the confidence we all ultimately seek. And so the reliance now is on machine learning based on sensor-provided data. This type of artificial intelligence approach extracts the information from data obtained by sensors, which is then offered to a machine learning program that identifies the signals that most discriminate based on experts’ subjective assessments. I am not aware of any study that has used an artificial intelligence approach with SCPE-generated clinical diagnoses. Perhaps appropriately so, given the unacceptably high rates of disagreement among experts. This leads to the inference that our only hope for reliable, exportable diagnostic information will have to be based almost entirely on objective (i.e. sensor-provided) data.