The proposed ND-PAE/FASD psychotropic algorithm predicted the paediatrician’s psychotropic medication choice for 54.1% of the clinic population, which suggests moderate accuracy according to Cohen’s Kappa statistic. While seemingly low, this incidence is promising for the future application of the algorithm. It should be noted that this algorithm’s intended use is alongside a clinician’s judgement. In this study, the prescription was completed without initial clinician guidance (a variance from the intended use of the algorithm) and being retrospective, this data collection missed the clinical context. In turn, the algorithm cannot account for changes in prescription due to side effects or ineffectiveness. Since the very purpose of the algorithm is to reduce polypharmacy, as expected, the algorithm failed to predict cases requiring polypharmacy (predicting single prescription cases (67.3%) vs multi prescription cases (27.8%)). This difference could be due to lack of clinician guidance or could be due to a limitation of the algorithm in the clinically complex cases.
Where the algorithm performed best was in single FASD cluster scenarios (67.3%), while it struggled in multi-prescription (27.8%) and no prescription cases (53.3%). We propose that the group differences are largely informed by the absence of clinician guidance. For no prescription cases, the algorithm recognises that non-psychotropic interventions are always first line for every cluster, however the algorithm cannot apply and evaluate such treatment by itself. The struggle with multi-prescription cases may further be informed by the variance of symptom presentation. A great challenge for treating FASD lies in the neurodevelopmental and symptom differences between individuals as well as within an individual throughout the developmental stages over a lifetime [8,9,10,11]. Moreover, the pediatric history is largely obtained second hand from a care giver. Future studies should examine if the accuracy of the algorithm increases with a pediatric patient’s age when individuals can give a self-report of symptoms. This may then suggest improved algorithm prediction if the patient can communicate on their own behalf.
We do recognise that the historical records are not infallible. The potential for misclassification of symptoms must be noted for FASD clusters with close overlap; in particular, the hyperarousal cluster with emotional dysregulation, as well as hyperactivity/neurocognitive cluster overlap with cognitive inflexibility. This overlap could inform some of the negative algorithm predictions. For instance, symptoms like hyperactivity and impulsivity can readily be misidentified as a behavioural disorder representing cognitive inflexibility, which in turn results in a negative prediction from the algorithm. As well, some researchers suggest ADHD may be over-diagnosed in children [16, 17]. This may explain why the algorithm over prescribed medication in some cases as the initial patient charts may have been influenced accordingly.
Some general patterns emerged from the algorithm. The algorithm often correctly predicted management for ND-PAE/FASD patients of the hyperactivity/neurocognitive cluster. We hypothesise this to be because externalizing behaviours are easier to identify by parents/caregivers. Within the cluster of hyperactivity/neurocognitive, the algorithm predicted medication for all cases of intellectual disability when this study data showed only half actually needed treatment (48%). This emphasises viewing the algorithm as merely a tool to be used alongside clinical judgement. Furthermore, the algorithm was unable to predict when some patients required more than one medication of a single class to treat their FASD cluster. We suggest that this is due to the spectrum of disease severity, which has already been noted in the literature .
With respect to cases that were removed for unclassified symptoms, the only cases removed were those where the symptom cluster could not be classified. If the patient had algorithm-classified symptoms amongst its unclassified ones but got a different medication than would be predicted by the algorithm, they were categorised as a “negative” case. Only cases where a conclusion could not readily be made were removed as unclassified. By looking at the medication prescribed by the physician in the cases that were removed, we inserted the symptom into the cluster most associated with that medication. Using this method, we propose an adjustment of the clusters as follows: sleep onset difficulty belongs in the hyperarousal cluster; gender dysphoria and obsessive compulsive disorder belong in the cognitive inflexibility cluster; grief belongs in the emotional regulation cluster; and autism spectrum disorder (ASD) belongs in the hyperactive/neurocognitive cluster (Table 3). Given the complex clinical presentation and the overlying executive dysfunction [18, 19], we would suggest the classification of ASD as hyperactive/neurocognitive. Future work should re-evaluate these suggested classifications, especially given our small unclassifiable sample size.
What reasons could explain why the algorithm did not predict the actual medication given in almost half of the cases? A reverse application of the algorithm upon actual clinical practise has some limitations. Judgement and guidance by the treating physician are needed to determine actual symptom severity, presence of impairment from these symptoms, relative impact of multiple symptoms which occur comorbidly, the mitigating effects of adjunct social supports and lifestyle adjustments (adequate sleep, and proper nutrition and exercise) and then, the final determination of whether or not any medication is warranted. Other factors affect choice, including parental preferences, priorities and medication affordability, and insurance coverage.
The results of this study are informed by one pediatrician’s treatment and impressions which affects the generalizability. Future work is needed for a matched comparison population with baseline-outcome findings. As the FASD algorithm research is still in infancy, we felt it best to utilise the whole sample size for a retrospective review.