Identifying epilepsy surgery referral candidates with natural language processing in an Australian context.

OBJECTIVE: Epilepsy surgery is known to be underutilized. Machine learning-natural language processing (ML-NLP) may be able to assist with identifying patients suitable for referral for epilepsy surgery evaluation. METHODS: Data were collected from two tertiary hospitals for patients seen in neurology outpatients for whom the diagnosis of "epilepsy" was mentioned. Individual case note review was undertaken to characterize the nature of the diagnoses discussed in these notes, and whether those with epilepsy fulfilled prespecified criteria for epilepsy surgery workup (namely focal drug refractory epilepsy without contraindications). ML-NLP algorithms were then developed using fivefold cross-validation on the first free-text clinic note for each patient to identify these criteria. RESULTS: There were 457 notes included in the study, of which 250 patients had epilepsy. There were 37 (14.8%) individuals who fulfilled the prespecified criteria for epilepsy surgery referral without described contraindications, 32 (12.8%) of whom were not referred for epilepsy surgical evaluation in the given clinic visit. In the prediction of suitability for epilepsy surgery workup using the prespecified criteria, the tested models performed similarly. For example, the random forest model returned an area under the receiver operator characteristic curve of 0.97 (95% confidence interval 0.93-1.0) for this task, sensitivity of 1.0, and specificity of 0.93. SIGNIFICANCE: This study has shown that there are patients in tertiary hospitals in South Australia who fulfill prespecified criteria for epilepsy surgery evaluation who may not have been referred for such evaluation. ML-NLP may assist with the identification of patients suitable for such referral. PLAIN LANGUAGE SUMMARY: Epilepsy surgery is a beneficial treatment for selected individuals with drug-resistant epilepsy. However, it is vastly underutilized. One reason for this underutilization is a lack of prompt referral of possible epilepsy surgery candidates to comprehensive epilepsy centers. Natural language processing, coupled with machine learning, may be able to identify possible epilepsy surgery candidates through the analysis of unstructured clinic notes. This study, conducted in two tertiary hospitals in South Australia, demonstrated that there are individuals who fulfill criteria for epilepsy surgery evaluation referral but have not yet been referred. Machine learning-natural language processing demonstrates promising results in assisting with the identification of such suitable candidates in Australia.

Delays in the diagnosis of ischaemic stroke presenting with persistent reduced level of consciousness: A systematic review.

INTRODUCTION: Stroke presenting with a reduced level of consciousness (RLOC) may result in diagnostic error and/or delay. Missed or delayed diagnosis of acute ischaemic stroke may preclude otherwise applicable hyperacute stroke interventions. The frequency, reasons for, and consequences of diagnostic error and delay due to RLOC are uncertain. METHOD: The databases PubMed, EMBASE, and Cochrane library were searched in adherence with the PRISMA guidelines. The systematic review was prospectively registered on PROSPERO. RESULTS: Initial searches returned 1162 results, of which 6 fulfilled inclusion criteria. The majority of identified studies show that ischaemic stroke presenting with RLOC is at increased risk of missed or delayed diagnosis. Hyperacute stroke interventions may also be delayed. There is limited evidence regarding the reason for these delays; however, the delays may result from neuroimaging delay associated with diagnostic uncertainty. There is also limited evidence regarding the outcomes of patients with stroke and RLOC who experience diagnostic delay; however, the available literature suggests that outcomes may be poor, including motor and cognitive impairment, as well as long-term impaired consciousness. The included studies did not evaluate, but have suggested urgent MRI access, educational interventions, and protocolisation of the evaluation of RLOC as means to reduce poor outcomes. CONCLUSIONS: Ischaemic stroke patients with RLOC are at risk of diagnostic delay and error. These patients may have poor outcomes. Additional research is required to identify the contributing factors more clearly and to provide amelioration strategies.

Identifying epilepsy surgery candidates with natural language processing: A systematic review.

INTRODUCTION: Epilepsy surgery is an underutilised, efficacious management strategy for selected individuals with drug-resistant epilepsy. Natural language processing (NLP) may aid in the identification of patients who are suitable to undergo evaluation for epilepsy surgery. The feasibility of this approach is yet to be determined. METHOD: In accordance with the PRISMA guidelines, a systematic review of the databases PubMed, EMBASE and Cochrane library was performed. This systematic review was prospectively registered on PROSPERO. RESULTS: 6 studies fulfilled inclusion criteria. The majority of included studies reported on datasets from only a single centre, with one study utilising data from two centres and one study six centres. The most commonly employed algorithms were support vector machines (5/6), with only one study utilising NLP strategies such as random forest models and gradient boosted machines. However, the results are promising, with all studies demonstrating moderate to high levels of performance in the identification of patients who may be suitable to undergo epilepsy surgery evaluation. Furthermore, multiple studies demonstrated that NLP could identify such patients 1-2 years prior to the treating clinicians instigating referral. However, no studies were identified that have evaluated the influence of implementing such algorithms on healthcare systems or patient outcomes. CONCLUSIONS: NLP is a promising approach to aid in the identification of patients that may be suitable to undergo epilepsy surgery evaluation. Further studies are required examining diverse datasets with additional analytical methodologies. Studies evaluating the impact of implementation of such algorithms would be beneficial.