Artificial intelligence in Emergency Medical Services dispatching: assessing the potential impact of an automatic speech recognition software on stroke detection taking the Capital Region of Denmark as case in point

The analysis suggests that a significant number of stroke calls are not detected as strokes (33.83%) within the 1-1-2 and 1813 emergency medical contact points. Considering the positive effects stroke recognition at the EMS takes on the stroke related outcome, the improvement of stroke detection at the EMS is crucial [14,15,16, 18, 19]. This research suggests the usage of an ASR, based on the model of CORTI AI for OHCA, to increase stroke recognition at the EMS from 52.75 to 61.19%. This increased detection rate through an ASR might decrease the number of multiple EMS calls for stroke patients, due to an earlier detection of the stroke and an accurate response within the first call. However, further research to determine the reason for multiple EMS calls would be necessary. Based on the condition that the stroke detection rate would increase by the same amount as the OHCA detection rate increased through CORTI AI, the rate of stroke patients treated with thrombolysis will rise by 5% within the group of stroke patients calling within time-to-treatment for thrombolysis [24, 54]. Additionally, the ASR might lead to an increase in thrombectomy of 8%, reperfusion of 8%, and surgical treatment of 2%. However, these increasing rates for thrombectomy, reperfusion and surgical treatment are to be viewed with caution. While the strength of the identified correlation between stroke recognition and treatment is moderate for thrombolysis, it is weak or very weak for the other treatment options. Additionally, the calculations have been made based on theoretical background and under the condition that the patients call the EMS within the treatment specific time-to-treatment. While 66.45% of all EMS contacts are within time-to-treatment of thrombolysis (4.5 h), 89.46% are within 24 h after stroke onset. Mosley et al. [55] confirm these findings by reporting that less than 50% of the stroke related calls were within 60 min after stroke onset [55]. In the future, a prospective study on the change in treatment through an increase in stroke detection would be interesting.

The described results suggest, that with an increased stroke detection at the EMS, the rate of stroke patients receiving endovascular treatment might decrease. The subgroup analysis of stroke patients with a time-to-call < 4.5 h showed a decrease in endovascular and surgical treatment. Nonetheless, this must be considered with caution, since endovascular treatment is regarded as an alternative for unsuccessful thrombolysis or patients not eligible for thrombolysis and surgical treatment is only carried out occasionally and under selected circumstances [56, 57]. While the time-to-treatment for thrombolysis is 4.5 h, endovascular treatment can be received within six to eight hours after stroke onset [56, 58]. Thus, patients who are not eligible for thrombolysis due to the closure of the window of time-to-treatment might receive endovascular treatment. In contrast, other reasons influence the choice of endovascular treatment [56]. Additionally, due to the low number of endovascular (n = 47) and surgical (n = 43) treatment, the results of these categories cannot be emphasized, but further research with a larger number of stroke patients treated with endovascular and surgical treatment would be necessary to draw conclusions [59].

Moreover, several additional factors, like stroke detection by the caller, recognition by the paramedic on scene, pre-conditions, and personal characteristics impact the stroke patients eligibility for treatment [14, 60]. Jones et al. [61] determined that symptoms like speech problems as well as posterior circulation symptoms were least likely to be recognised as stroke related. Further research on the beforementioned connections as well as on mortality and on the score of the modified ranking scale, which defines a patients clinically discrete disability caused by a stroke on a scale of seven levels, would be helpful in order to draw precise and grounded conclusions on the effect of EMS stroke detection on patient outcome [62, 63]. Nonetheless, stroke detection by the EMS might impact the treatment, specifically thrombolysis. The relevance of an ASR for stroke detection at the EMS is underlined by the substantial amount (49% of “stroke relevant criteria but no “A” response” and 60.93% of “stroke nonrelevant criteria”) of calls within time-to-treatment for thrombolysis in the categories “strokes not detected”.

Based on the results of the analysis, it can be argued that the ASR could specifically impact the detection of those characteristics with a negative correlation to “stroke detected” or a positive correlation to one of the categories within “strokes not detected”.

The analysis indicates that an improvement of stroke detection is particularly important for calls to the 1813 Medical Helpline, due to the observed negative correlation of stroke detection within 1813-calls. Thus, the ASR should be used for both access numbers 1813 and 1-1-2. The negative correlation may be influenced by non-recognition of atypical stroke symptoms by the caller, thus the 1813 instead of the 1-1-2 is called [64]. However, for validation further research is needed.

When training the ASR specific attention should be placed on haemorrhagic strokes, due to the positive correlation between haemorrhagic strokes and “stroke nonrelevant criteria” and the small representation of haemorrhagic strokes (9.22% of all strokes). Several authors argue, that it is particularly important to take into account underrepresented groups, e.g. haemorrhagic stroke patients (n = 834) and patients with “stroke nonrelevant criteria” (n = 1215), when training an ASR, in order to avoid a bias, that could possibly cause an erroneous stroke detection algorithm [65, 66]. An ASR could also positively influence the stroke detection rate of females, due to the negative correlation to stroke detection. Lisabeth et al. [67] and Rathore et al. [68] support this finding by describing, that women reported a larger amount of non-traditional stroke symptoms.

In the data analysis a negative correlation between stroke detection and weekends was determined, hence the ASR for stroke detection could particularly improve the stroke detection on weekends. A possible explanation could be, that on weekends 53.11% of all stroke related EMS calls are to the 1813, while within the week only 37.1% are to the 1813. As argued before, 1813-calls might entail more atypical stroke symptoms not detected compared to the 1-1-2, resulting in a decline of the detection rate on the weekends [64].

Interestingly, stroke patients within the group “stroke relevant criteria but no “A” response” are significantly younger than the patients within the other groups. Considering the research by Singhal et al. [69], detection of stroke among younger patients, is challenging due to infrequency in comparison to stroke mimics and missing awareness among the general population as well as the EMDs. This might result in an EMS contact outside of the window of time-to-treatment or missing recognition of severity and thus in no “A” response. This is strengthened by “stroke relevant criteria but no “A” response” having the lowest proportion of calls (49%) within the time-to-treatment for thrombolysis. This reasoning is also supported by the subgroup analysis showing no statistically significant difference in age between the categories “stroke detection” and “strokes not detected”.

The category “stroke relevant criteria but no “A” response”, has a different distribution throughout the time of day, compared to strokes detected and strokes with non-relevant criteria. While the latter have a peak between 8 a.m. and 10 a.m. and thereafter steadily decrease, the beforenamed category is comparatively steady between 8 a.m. and 6 p.m. The peak of stroke relevant criteria in the morning might be due to the so called “wake-up stroke”, for which EMDs have a high awareness, since one out of five strokes is a “wake-up stroke” [70]. Comparably, in the afternoon a greater diversity among emergency calls occurs, which might result in a higher difficulty to detect strokes [71]. For these calls an ASR supporting the EMD in the stroke detection would be useful to detect and send the correct response. Due to the subgroup analysis identifying no statistically significant difference in time of day between “stroke detection” and “strokes not detected” an ASR would be relevant for increasing stroke detection throughout the whole day. The steady number of calls with “stroke relevant criteria but no “A” response” might be caused by a delayed emergency call and despite the stroke detection by the dispatcher, but due to the closure of the window of time-to-treatment, no “A” response. This argument is supported by the outcome of the subgroup analysis, showing no statistically significant difference in time of day between “stroke detection” and “stroke relevant criteria but no “A” response”. Further research to conduct the reason for stroke relevant criteria but no “A” response is necessary. The “missing criteria” show two peaks throughout the day, between 8 a.m. and 10 a.m. as well as between 4 p.m. and 6 p.m. These peaks can be explained by the majority of “missing criteria” within 1813 calls, and the increased amount of 1813 calls between 8 a.m. and 10 a.m. on the weekends and between 4 p.m. and 6 p.m. during the week, due to its mission as out-of-hours general practitioner [13]. However, additional factors that might be influenced by an ASR, were not considered in this study.

Due to a significant increase of stroke detection throughout the years 2016–2018, as shown in our analysis, it might be argued that no further technical support might be necessary to improve stroke detection. However, since a significant decrease has only been seen within the group “stroke relevant criteria but no “A” response”, but not within the group “stroke nonrelevant criteria”, this argument can be discarded due to the ASR presumably impacting the recognition of strokes with currently “stroke nonrelevant criteria”, by increasing the detection of stroke symptoms and thus indicating “stroke relevant criteria”. Additionally, the subgroup analysis indicating no statistically significant increase in stroke detection throughout the years 2016–2018 supports the need of an ASR for improving stroke recognition by the EMDs. The increase in stroke detection seen for 2016–2018 might be influenced by the publication by Viereck et al. [15] in 2016 on the recognition of strokes through EMDs, after which small changes have been made in the algorithm of the 1-1-2. Another reason for the change in stroke detection throughout the years 2016–2018, could be the results of a research conducted at the University of Kentucky Stroke Center impacting the stroke recognition campaign, “FAST” (Face, Arm, Speech, Time) to “BE-FAST” (Balance, Eyes, Face, Arm, Speech, Time) in 2017, through including visual symptoms on stroke [72]. This revision might have led to an increasing sensibility for strokes within the population possibly resulting in a clearer expression of the symptoms to the EMS and an increasing sensibility of EMDs for stroke related symptoms [72].

The question arises, whether other options could increase stroke detection by EMS call-takers. Past research analysed the influence of educational training modules as well as stroke recognition scales and protocols, such as the “FAST”-Tool [17, 73,74,75]. However, Oostema et al. [73] reported, that the increase in stroke recognition after an educational intervention was limited to three months and might increase the rate of false positive stroke detection due to a higher sensibility to symptoms related to stroke [73, 76]. Additionally, the systematic review by Oostema et al. [17] discovered, that the correct usage of the scales and protocols has not been analysed in the included studies, resulting in lacking security of the right usage. It is to be mentioned, that educational programmes for EMDs might increase the rate of false positive stroke detection due to a higher sensibility to symptoms related to stroke [76].

Like the correct use and acceptance of scales and protocols, the acceptance and adoption of the ASR into the EMS call by the EMD, is relevant for its effect on stroke detection. Blomberg et al. [77] reported a lack of compliance with the suggestions of CORTI AI by the EMDs, which resulted in no increase of OHCA detection within the EMS Copenhagen. Considering the results of educational interventions, the introduction of an ASR for strokes at the EMS could be accompanied by, for example educational interventions addressing challenges in the uptake of the ASR, in order to ensure the effect of the ASR [73,74,75, 77]. The European Institute of Innovation and Technology (EIT) Health states that to improve the uptake and effect of AI in healthcare, investments in the education of healthcare workers to ensure digital literacy, the exchange of best practice in the field of AI in healthcare throughout the EU and improvement of collaboration is essential [78].

Despite the lack in compliance with the ASR and thus the limitation of the effect, no sole usage of an ASR should be aimed for, due to possible input and algorithm bias as well as the missing consideration of the emotional component [79, 80]. In summary, the combination of an ASR with a well-trained human professional can substantially increase the number of correctly detected strokes [24, 25].

Limitations

The definition of stroke detection as “stroke relevant criteria” and an “A” response, might not represent all the strokes detected within the EMS. Possibly, strokes were detected within the category “stroke relevant criteria but no “A” response”, and still, due to the closure of the window of time-to-treatment no “A” response was sent. For those cases obviously, an ASR would not impact the stroke detection. Contrarily, strokes might have been detected within the category “missing criteria”, but no criteria were indicated within the system, yet an “A” response had been sent as the correct stroke response. Likewise, possibly “stroke nonrelevant” criteria were indicated within the system, but the EMD recognised the stroke and sent an “A” response. Due to the definition of stroke response within the EMS Copenhagen, the proxy of “stroke relevant criteria” and “A” response was considered the most accurate to define stroke detection for this research.

Another limitation of the stroke related emergency calls is, that all EMS calls seven days prior and post stroke were included within this study, even if the emergency call was not related to the stroke of the patient but was due to another medical issue. However, research has shown that strokes typically impact the health of the patient significantly, through post-stroke and pre-stroke symptoms, thus the number EMS contacts of stroke patients not related to the stroke might be comparably small [36, 67]. The choice to include stroke calls seven days prior and post stroke could be affecting the response made by the medical dispatcher, depending on the time of symptom onset named within the call and thereby diminish the effect of the outcome. Unfortunately, the data on time of symptom onset is not documented and thus not available and must therefore be considered a blind spot within this research. Additionally, the subgroup and time-to-call analysis is limited due to the determination of stroke onset within the patient-doctor consultation based on the patients recall of time of symptom onset. Thus, the possibility of recall bias needs to be considered in the interpretation of the results [81, 82].

The internal validity, which is described as to which extent the study accurately measures the concept, might be also limited due to the assumption, that an ASR for stroke has the same effect on the increase of detection at the EMS, as CORTI AI on OHCA, since OHCA symptoms are more specific compared to stroke symptoms [33, 83,84,85,86]. Thus, the possibility of stroke mimics, which are defined as disorders showing stroke symptoms, such as for example brain tumours, metabolic disorders, or migraines, and are diagnosed as strokes are likelier than false positive OHCA [85]. This is supported by the research by Watkins et al. [84] detecting a specificity of 99.4% within OHCA, while according to Hatzitolios et al. [85] 5% and to Hosseininezhad and Sohrabnejad [86] 14.9% of all stroke-like symptoms are stroke mimics. However, since CORTI AI for OHCA is, to the researcher’s knowledge, the only ASR within an EMS context, the presumable increase of 16% based on CORTI AI was chosen. Hence, this limitation must be considered when referring to the presumable increase in stroke detection, especially since Blomberg et al. [24] reported a decrease in specificity for OHCA detection with the ASR from 98.8 to 97.3% (p < 0.001). Under consideration of the beforenamed rate of stroke mimics, the decrease in specificity of stroke detection might be higher with an ASR compared to the decrease in specificity of OHCA. Further research on the topic of specificity in stroke detection through ASR should be performed to elaborately address this point and to discuss possible mitigation strategies.

The transferability to the population of Denmark would need further research, since the results conducted for the Capital Region of Denmark, with the specialty of the 1813, might not be transferable to the entirety of Denmark [87]. Additionally, the transferability to other countries might be limited, due to country specific EMS and population characteristics. Thus, the assessment of transferability on superordinate level using for example the PIET-T Model might be helpful [88]. Because of the mentioned limiting factors, the results of this study should be interpreted with caution and considered as directing and indicating further research fields.

留言 (0)

沒有登入
gif