Abstract
Introduction: Erroneous penicillin allergy labels are associated with significant health and economic costs. This study aimed to determine whether deep learning-facilitated proactive consultation to facilitate delabelling may further enhance inpatient penicillin allergy delabelling. Methods: This prospective implementation study utilised a deep learning-guided proactive consultation service, which utilised an inpatient penicillin allergy delabelling protocol. The intervention group comprised all admitted inpatients with a penicillin allergy over the course of a 14-week period in a tertiary hospital. The rate of penicillin allergy delabelling in the intervention group was compared to that of a historical control group. Results: There were 439 patients included in the study, of whom 121 were identified by the algorithm as suitable for penicillin allergy interrogation. Of those identified by the algorithm, 16.5% were successfully delabelled in the inpatient setting within the same admission and 9.9% were referred for outpatient testing. This result was statistically significantly greater compared to the rate of delabelling in the historical control group (0%, p = 0.00001). There were no adverse reactions. The projected annual savings associated with the program over a 12-month period were AUD 1,170,617.16. Conclusion: Deep learning-facilitated proactive inpatient penicillin allergy delabelling was effective, safe, and economical in this single-centre implementation study. Further studies should seek to examine this approach in diverse centres.
Introduction
Penicillin allergy labels are associated with significant health and economic costs [1, 2]. Patients with these labels are prescribed alternative antibiotics which may be less effective, more costly [2]. As a result, these patients have increased length of stay, hospital readmission rate, and post-operative complications compared to those without penicillin allergy labels [3‒5]. The alternative non-penicillin antibiotics often have unnecessarily broad antimicrobial coverage, increasing the risks of antibiotic resistance and Clostridium difficile infection [4].
It is estimated that approximately 10% of inpatients carry a penicillin allergy label [4, 6]. However, when tested, 90% of these patients are found to tolerate penicillin and can be safely “delabelled” [7]. Delabelling patients of their penicillin allergy label has been shown to be cost-effective when conducted opportunistically in the inpatient setting, leading to cost savings through reduction in follow-up consultations, opportunity cost, and car travel costs for patients [8, 9]. However, currently, the majority of penicillin allergy delabelling is carried out by specialist immunologists in the outpatient setting [1]. Penicillin allergy labels are rarely evaluated on suitability for delabelling during inpatient admissions, and as a result, these patients are rarely referred to the immunology service for consideration of this procedure [1]. Such referrals are often initiated by the treating team on a “reactive” basis when prompted by a lack of alternative antibiotic options available [1].
Despite the recent exponential growth of artificial intelligence-based technological applications across various fields of medicine, particularly in radiology, cardiology, and haematology, there are currently no registered artificial intelligence-based applications being utilised in clinical practice within the field of allergy, as of February 2023 [10]. Natural language processing is one branch of artificial intelligence that facilitates the evaluation of free-text notes in electronic medical records (EMRs) into interpretable datasets, which can aid in prediction or triage [11]. One such example includes the successful application of natural language processing to predict hospital admission of patients in the emergency department, which enables augmentation of the existing triage process [12]. In the 2022 American Academy of Allergy, Asthma, and Immunology working group report, one suggested application of natural language processing was to enter clinical and demographic parameters as “inputs” into a neural network, to predict the risk of penicillin allergy in patients [13].
This study investigates the use of a validated machine learning algorithm to identify inpatients suitable for proactive penicillin allergy delabelling. This subset of patients identified by the algorithm are offered access to an opportunistic penicillin allergy delabelling program. The aim of the study was to determine if machine learning-facilitated oral penicillin allergy delabelling program will reduce the number of individuals with incorrectly labelled penicillin allergies.
Methods
This is a single-centre prospective study, conducted in an 800-bed metropolitan tertiary referral adult general hospital, Royal Adelaide Hospital (RAH). The EMR utilised was the Allscripts® Sunrise™ Enterprise Release 17.3 [14] which has a dedicated section for adverse reaction labels (online suppl. Information 1; for all online suppl. material, see https://doi.org/10.1159/000542589).
Machine Learning Algorithm
Patients in both the intervention group and historical control group [15] were identified by applying the previously derived machine learning algorithms, to systematically identify RAH inpatients with penicillin allergy labels who may be suitable for penicillin allergy evaluation. Derivation and validation of the algorithm was previously described in our earlier published studies. In the first study, machine learning natural language processing was used to analyse free-text penicillin adverse reaction data extracted from the EMR [16]. In this study, there were 3,963 penicillin adverse reaction labels with sufficient information to enable classification [16]. The adverse reaction entries were split into training and testing datasets (75%/25% split) to develop and test a variety of different machine learning models including artificial neural networks, logistic regression models, random forest models, and Naïve Bayes models [16]. The six-layer artificial neural network was the best performing model for the classification of allergy versus intolerance (AUC 0.994, sensitivity 0.99, specificity 0.96), and also for classification of high or low risk of true allergy (AUC 0.988, sensitivity 0.99, specificity 0.99) [16]. There were six fully connected layers; the numbers of nodes were 500, 200, 100, 100, 10, and 4, respectively, with a final output layer with one node [16]. This was the model employed in the present study.
The algorithm classifies documented adverse reactions through text pre-processing, involving negation detection, word stemming, and stopword removal [16]. These classifications were utilised as previously defined using expert criteria [16‒18]. The expert criteria were developed in consultation with an expert panel consisting of an immunologist and clinical pharmacologist, in order to classify (1) allergy versus intolerance, as well as (2) risk of penicillin allergy based on the documented reactions [9]. To minimise false-positive delabelling-target results, the study implemented conservative algorithm thresholds.
In the second study, the machine learning models were subsequently externally validated on a temporally separated dataset of adverse reactions from five public hospitals in South Australia [19]. The same artificial neural network architecture and weights were used again. The models performed comparably with expert criteria in the categorisation of allergy versus intolerance, and identification of high-risk versus low-risk allergies [19].
Participant Recruitment
All individuals included in the study were over 18 years of age, admitted on a screening day (weekly) under any service during the study period at the RAH, and with an allergy label in the EMR to one of the several nominated penicillin antibiotics identified by generic medication names. These antibiotics included “penicillin,” “amoxicillin,” “amoxicillin-clavulanate,” “flucloxacillin,” “piperacillin,” and “piperacillin-tazobactam.” Individuals in the control group were identified in a historical 20-week study period. Individuals in the intervention group were identified in a subsequent 14-week study period.
Patients in the intervention group were approached in the inpatient setting by a medical officer from the immunology service and were offered the opportunity to participate in the oral penicillin allergy delabelling program. The oral penicillin allergy delabelling program involves inpatient evaluation using a structured penicillin allergy history, and the provision of options for penicillin allergy delabelling in either the inpatient or outpatient setting. The criteria for inclusion in the oral penicillin allergy delabelling program were (1) current inpatient at the RAH who is documented as having an allergic reaction to penicillin in the EMR, (2) ≥18 years, and capable of providing informed consent. The exclusion criteria included (1) patients deemed by the treating team to be too unstable or as having another clinical contraindication for participation (e.g., end of life), (2) pregnant or breastfeeding patients, (3) patients in whom prescription of penicillin will interfere with clinical care, e.g., those where the treating team are trying to culture an organism and administering penicillin will interfere with this.
A historical control group from a previously published study [15] in the same centre was used as a comparator for this study. All patients in the control group were also identified by the same machine learning algorithm as having a penicillin adverse reaction label potentially amenable for delabelling. The purpose of the historical control group was to determine the baseline rate of inpatient penicillin allergy delabelling at the RAH, prior to implementation of the proactive delabelling program. The delabelling rate in the historical control group represents the current standard of care – patients are still able to access inpatient penicillin allergy delabelling via the inpatient immunology consult service; however, this service was only available upon request by the inpatient treating team. On the other hand, patients in the intervention group were proactively offered access to an opportunistic oral penicillin allergy delabelling program.
Oral Penicillin Allergy Delabelling Program
For all patients consenting to further in-person allergy evaluation in the oral penicillin allergy delabelling program, a structured penicillin allergy history was undertaken by an immunology medical officer using a standardised form (online suppl. Information 2). The collected information was documented in the EMR. Based upon the structured allergy history, the validated beta-lactam antibiotic allergy assessment tool (AAAT) (2) was applied to facilitate assignment of a standardised (i) phenotype and (ii) management option, depending on the clinical manifestations of the index reaction. The phenotypes of the AAAT included (i) severe immediate hypersensitivity (IgE mediated), (ii) non-severe immediate hypersensitivity (IgE mediated), (iii) severe delayed hypersensitivity (T cell mediated), (iv) non-severe delayed hypersensitivity (T cell mediated), (v) potential immune mediated (e.g., acute interstitial nephritis), or (vi) unlikely to be significant/non-immune mediated (e.g., gastrointestinal upset, unknown history). (2) In cases of uncertainty, individual cases were discussed with an immunology specialist to determine the most appropriate management option.
All patients enrolled in the oral penicillin allergy delabelling program received individualised education on the significance and implications of their penicillin allergy label and were offered options to facilitate delabelling, as well as the option to decline further testing. Depending on the clinical phenotype, one of the following management options was undertaken: (i) direct delabelling with removal from medical record, (ii) oral challenge (provocation testing), (iii) antibiotic allergy skin testing followed by oral challenge, or (iv) outpatient antibiotic allergy assessment involving detailed specialist consultation with further testing thereafter as required. For patients suitable for an oral challenge, the option was provided for this to be undertaken in either the inpatient or outpatient setting, depending on the patients’ clinical stability as well as the preference of the patient and their treating team. The oral challenge was preferentially undertaken in the inpatient setting where possible.
Clinical Testing Protocol
For patients who were deemed “appropriate for supervised direct oral rechallenge” according to the AAAT score, or as determined by specialist immunologist consultation, a two-step graded oral amoxicillin challenge was undertaken. On day 1, 50 mg oral amoxicillin liquid was administered, followed by 1 h of observation. If no reactions occurred, 500 mg oral amoxicillin was administered followed by 2 h of observation. The following day, the patient commenced a 3-day course of 500 mg oral amoxicillin 8 hourly.
Patients who were deemed to be of minimal risk and were suitable for direct delabelling according to the AAAT score, but were unwilling to be delabelled without a challenge, received a single dose of 500 mg oral amoxicillin, followed by 2 h of observation. Patients deemed “appropriate for skin testing followed by oral rechallenge” according to the AAAT, or as determined by specialist immunologist consultation, were referred for outpatient skin prick testing and intradermal testing in the medical day unit (online suppl. Information 3). Follow-up protocols and statistical analyses, with cost estimates based on previous research [8], are outlined in online supplementary Information 4.
Results
Patient Demographics
There were 439 patients included in this study. In the intervention group, there were 121 individuals identified by the algorithm during the study period as suitable for penicillin allergy evaluation. The demographics are described in Table 1. The median age was 74.59 years (interquartile range 63.4–83.6). There were 75 females (62.0%). In the control group [15], there were 318 patients, with a median age of 68.7 years (IQR 54.2–81.3) and 175 (55.0%) were female. The patients in the intervention group were admitted under a variety of services including general medicine (34.8%), orthopaedic surgery (10.9%), and cardiology (10.9%).
Patient demographics
. | Intervention group . | Control group . |
---|---|---|
Total number of patients | 121 | 318 |
Female, n (%) | 24 (52.2) | 175 (55.0) |
Age (median, interquartile range) | 74.59 (63.4–86.3) | 68.7 (54.2–81.3) |
. | Intervention group . | Control group . |
---|---|---|
Total number of patients | 121 | 318 |
Female, n (%) | 24 (52.2) | 175 (55.0) |
Age (median, interquartile range) | 74.59 (63.4–86.3) | 68.7 (54.2–81.3) |
In total, of the 121 patients identified by the algorithm, 16.5% were successfully delabelled in the inpatient setting within the same admission. This represents a statistically significantly higher rate of inpatient penicillin allergy delabelling compared to the control group (0%, p = 0.00001). Additionally, 9.9% of patients identified by the algorithm were referred for outpatient testing.
Of the 46 patients enrolled in the oral penicillin allergy delabelling program, 23.9% (11/46) were delabelled through direct delabelling without testing, 19.6% (9/46) were delabelled through inpatient challenge, and 26.1% were referred for outpatient testing. Additionally, 11 patients enrolled in the program did not wish to undergo further evaluation (23.9%), 1 patient was referred to the immunology outpatient clinic, and 1 patient’s penicillin allergy label was revised. The predominant reason provided for declining further evaluation was due to preoccupation with current medical and surgical issues, including the new diagnosis of a life-limiting illness. Overall, the rate of inpatient delabelling in the intervention group was 16.5% which was statistically significantly higher than in the historical control group (0%, p = 0.00001).
The drug culprits listed on the adverse reaction label were (Table 2) penicillin (n = 36, 78.3%), amoxicillin (n = 5, 10.9%), amoxicillin-clavulanate (n = 3, 6.5%), flucloxacillin (n = 1, 2.2%), ampicillin (n = 1, 2.2%). Among the individuals included in the delabelling program, the “reaction” section on the allergy label described “rash” in 27 patients and “rash; itching” in 1 patient. There were 5 patients who had an adverse reaction label categorised by the user as an “allergy” (as opposed to an intolerance), despite the listed reactions being indicative of gastrointestinal intolerance, labelled as “diarrhoea” (n = 2), “GI upset; abdominal pain” (n = 1), “nausea” (n = 1), “nausea/vomiting” (n = 2).
Culprit agents and phenotypic descriptions of index reactions for patients in the intervention group
Culprit agents listed on the adverse reaction label . | N = 46 . |
---|---|
Penicillin | 36 (78.3%) |
Amoxicillin | 5 (10.9%) |
Amoxicillin-clavulanic acid | 3 (6.5%) |
Flucloxacillin | 1 (2.2%) |
Ampicillin | 1 (2.2%) |
Culprit agents listed on the adverse reaction label . | N = 46 . |
---|---|
Penicillin | 36 (78.3%) |
Amoxicillin | 5 (10.9%) |
Amoxicillin-clavulanic acid | 3 (6.5%) |
Flucloxacillin | 1 (2.2%) |
Ampicillin | 1 (2.2%) |
Reaction description (according to AAAT classifications) . | N = 46 . |
---|---|
Diffuse rash or localised rash with no other symptoms, >24 h after starting antibiotic, reaction occurred >10 years ago | 15 (32.6%) |
Gastrointestinal symptoms | 6 (13.0%) |
Unknown reaction >10 years ago | 5 (10.9%) |
Unknown reaction (and unknown time period since reaction) | 2 (4.3%) |
Childhood exanthem | 6 (13.0%) |
Swelling | 1 (2.2%) |
Neurological or CNS manifestation | 2 (4.3%) |
Diffuse rash or localised rash with no other symptoms >24 h post starting antibiotic <10 years ago | 2 (4.3%) |
Angioedema | 1 (2.2%) |
Urticaria | 1 (2.2%) |
Eosinophilia | 1 (2.2%) |
Other reaction descriptions, unable to be classified using AAAT | |
Pruritus without rash | 1 (2.2%) |
Immediate localised rash around abdomen with both amoxicillin and cefalexin | 1 (2.2%) |
Immediate localised rash localised to the chest | 1 (2.2%) |
Vaginal thrush | 1 (2.2%) |
Reaction description (according to AAAT classifications) . | N = 46 . |
---|---|
Diffuse rash or localised rash with no other symptoms, >24 h after starting antibiotic, reaction occurred >10 years ago | 15 (32.6%) |
Gastrointestinal symptoms | 6 (13.0%) |
Unknown reaction >10 years ago | 5 (10.9%) |
Unknown reaction (and unknown time period since reaction) | 2 (4.3%) |
Childhood exanthem | 6 (13.0%) |
Swelling | 1 (2.2%) |
Neurological or CNS manifestation | 2 (4.3%) |
Diffuse rash or localised rash with no other symptoms >24 h post starting antibiotic <10 years ago | 2 (4.3%) |
Angioedema | 1 (2.2%) |
Urticaria | 1 (2.2%) |
Eosinophilia | 1 (2.2%) |
Other reaction descriptions, unable to be classified using AAAT | |
Pruritus without rash | 1 (2.2%) |
Immediate localised rash around abdomen with both amoxicillin and cefalexin | 1 (2.2%) |
Immediate localised rash localised to the chest | 1 (2.2%) |
Vaginal thrush | 1 (2.2%) |
AAAT, antibiotic allergy assessment tool.
Details of the index reactions to penicillin were collected using the structured penicillin allergy history form and categorised using the AAAT. For the patients in the oral penicillin allergy delabelling program, reported reactions were categorised based upon phenotype (Table 2). Additionally, there were 4 patients that reported index reactions which could not be classified using the AAAT, including “pruritis without rash” (n = 1), “immediate localised rash around abdomen with both amoxicillin and cefalexin” (n = 1), “immediate localised rash localised to the chest” (n = 1), and “vaginal thrush” (n = 1).
Directly Delabelled
Of the 46 patients enrolled in the oral penicillin allergy delabelling program, 11 (23.4%) were directly delabelled. For these patients, the allergy label was removed based upon the structured allergy history, without the need for formal testing. The AAAT clinical manifestation based on the structured allergy history undertaken for these patients was consistent with “gastrointestinal symptoms” for 6 of these patients, including symptoms of nausea, vomiting, abdominal pain, or diarrhoea. In all six cases, the patients’ documented reaction on their allergy label (e.g., “GI Upset; Abdominal pain”) was consistent with the findings of the structured allergy history. One (2.1%) patient reported a history of auditory hallucinations, corresponding to the low-risk AAAT category of “Neurological manifestation” which was directly delabelled. One (2.1%) patient had a documented allergy history labelled with “rash” as the reaction; however, structured allergy history revealed the index reaction was consistent with vaginal thrush and thus was also directly delabelled. This patient subsequently tolerated amoxicillin.
During the study period, four (8.5%) inpatients had inadvertently received multiple doses of a penicillin-based antibiotic during the present admission, prior to formal allergy assessment; however, their penicillin allergy label had not been removed from the EMR. Structured allergy assessment determined that none of the patients developed an adverse reaction to the inadvertent administration of penicillin; therefore, the penicillin allergy labels were delabelled without further intervention.
Inpatient Oral Amoxicillin Challenge
The number of patients who underwent a penicillin challenge was 9 (19.6%). All challenges undertaken were successful and led to delabelling. No immediate-type or delayed-type adverse reactions occurred. Five patients received a two-step oral amoxicillin challenge, and 3 patients received a one-step oral amoxicillin challenge. One patient completed a 3-day course of intravenous piperacillin-tazobactam, under monitoring, as an inpatient, due to the presence of a therapeutic indication for this agent in the admission. This patient subsequently completed a one-step oral amoxicillin challenge to facilitate delabelling of an amoxicillin allergy. Routine patient follow-up conducted in person or via telephonic communication at approximately the 7-day mark confirmed that no delayed-type adverse reactions occurred. All patients answered “yes” when asked if they were willing to use penicillins in the future.
Outpatient Evaluation
The number of patients referred for outpatient testing was 12 (25.5%). Of these patients, ten were referred for a two-step oral amoxicillin challenge, and two were referred for SPT, IDT, and oral amoxicillin challenge. Additionally, 1 patient had been referred for formal outpatient evaluation with detailed specialist consultation at the request of the treating team.
Other
The penicillin allergy label for 1 patient was updated to a less restrictive label: the drug culprit was modified from “penicillin” to “piperacillin-tazobactam,” and the reaction was revised to permit the use of amoxicillin in future. One patient agreed to participate in the delabelling program but was determined not suitable for further testing due to concurrent participation in a clinical trial.
Cost-Saving Estimate
A recent systematic review reported that penicillin allergy delabelling resulted in antibiotic cost savings at a (1) patient level of AUD 225–7,800 per delabelled patient and (2) annual hospital drug savings between AUD 12,400 and 26,000 [20]. In the Australian context, a more recent health economic study provided a cost comparison of inpatient admission pre- and post-penicillin allergy delabelling program and estimated that access to the delabelling program resulted in a mean difference of AUD –9,467.72 (95% confidence interval [CI]: AUD –15,419.98, AUD –3,515.46; p = 0.002) per person [8]. This was calculated as a cost comparison of inpatient admissions pre- and post-penicillin allergy delabelling program [8]. Assuming the benefits are generalisable to the intervention in this study, if all patients delabelled (or referred for delabelling) in the 3-month study period were to have one further inpatient admission in their lifetime, the projected cost savings of our intervention would amount to approximately AUD 312,434.76 [8]. If this was extrapolated to a 12-month time period, the estimated projected cost savings would be AUD 1,211,868.16. In this study, the mean cost of the penicillin allergy delabelling intervention per effectively delabelled patient was AUD 20.51 in the inpatient hospital setting, and AUD 362.48 in the outpatient setting. Using these published figures, the cost of our delabelling program in the inpatient and outpatient settings over a 3-month period was estimated to be AUD 4,780.47. In addition to this, allowing for a 0.5 full-time equivalent position to complete the delabelling, the overall cost of running the intervention over 3 months was approximately AUD 19,780.47. Therefore, the projected net savings of implementing our delabelling program for 3 months could amount to AUD 292,654.29, which over 12 months would total AUD 1,170,617.16.
Discussion
This study has successfully shown the use of artificial intelligence facilitating a novel clinical workflow. In this study, the artificial intelligence algorithm provided utility essentially through a process of triaging. The implementation of a specialist service that proactively identifies and triages inpatients with penicillin allergy labels on an opportunistic basis is one strategy that can be used to increase the rate of penicillin allergy delabelling. However, in an 800-bed hospital with high inpatient turnover, manual screening of the EMR to identify and risk-stratify all penicillin allergy labels to identify potential delabelling targets is a time-intensive activity and, therefore, challenging to feasibly implement. In this study, deep learning was successfully used to facilitate the efficient identification and risk stratification of patients with penicillin allergy labels who were deemed likely amenable to delabelling.
In our study, we found that a high proportion (92.0%, 46/50) of patients identified by the machine learning algorithm and who were eligible for enrolment (i.e., those who had not been discharged prior to the first opportunity to review and who did not meet exclusion criteria) were enrolled in the oral penicillin allergy delabelling program, and 40% (20/50) were successfully delabelled as inpatients. Of the patients offered the opportunity to enrol in the oral penicillin allergy delabelling program, 96.6% agreed to enrol, indicating a high level of patient acceptability of the program. All inpatient treating teams were contacted prior to patient enrolment, and 100% of teams agreed to enrolment of their patient in the oral penicillin allergy delabelling program, indicating a high level of clinician acceptability. This acceptability, while in part related to penicillin allergy delabelling, also shows that artificial intelligence screening can identify patients for studies who would be keen to participate, who would otherwise be missed. There was a statistically significantly higher rate (p = 0.0001) of inpatient penicillin allergy delabelling (16.5%) in the intervention group (offered access to the proactive penicillin allergy delabelling program), compared to patients with penicillin allergy labels in the control group (0% delabelled) who had access to the standard inpatient immunology consult service as part of routine care but were not offered access to the proactive penicillin allergy delabelling program. This suggests that the proactive penicillin allergy delabelling program significantly improves the rate of delabelling compared to current clinical practice, which relies on delabelling on an “as-needed” basis.
This study provides a model for future deep learning-assisted proactive delabelling interventions. Similar approaches could be employed for future interventions, such as with respect to diabetes and blood glucose level optimisation, antimicrobial stewardship, or anticoagulation for atrial fibrillation.
Opportunistic identification of potential penicillin allergy delabelling targets in the inpatient setting, as demonstrated in this study, is particularly advantageous, as it has previously been demonstrated that penicillin allergy delabelling programs are significantly more cost-effective when undertaken in the inpatient setting (AUD 20.51 per effectively delabelled patient) compared to the outpatient setting (AUD 362 per effectively delabelled patient) [8]. In our institution, arranging an outpatient oral challenge necessitates multiple patient interactions and “visits.” This involves an initial clinic appointment to facilitate history-taking, as well as a pre-admission phone consultation, a day unit admission for the challenge itself, and a follow-up phone consultation to discuss the results. An opportunistic inpatient approach negates the need for these additional outpatient visits, eliminates additional travel costs, reduces time lost for the patient and carer, avoids parking costs, and reduces direct health service costs in terms of reduced nursing and medical staffing requirements [8].
The projected cost savings associated with this study were substantial and likely would be a self-sustainable service from a funding perspective due to improvements in future healthcare utilisation. Having a penicillin “allergy” label has previously been found to be associated with an increased length of stay, with an average of 0.59 (9.9%; 95% CI, 0.47–0.71) more total hospital days during a 20.1 ± 10.5 month follow-up period in a US study [4]. In an Australian study, the mean difference in length of stay was estimated to be −4.41 days (95% CI: −7.64, −1.18) between the delabelled patients compared with the non-delabelled usual care cohort [8].
This study has limitations, including its single-centre status. Another limitation is the use of a historical control cohort with lack of randomisation. Future research should aim to substantiate the preliminary findings from our study using a multi-centred study with longitudinal follow-up to determine rate of relabelling.
Conclusion
Machine learning was successfully used to facilitate the safe and efficient implementation of an opportunistic, inpatient penicillin allergy delabelling program. This proactive, systematic approach of penicillin allergy delabelling significantly reduced the number of inpatients with incorrectly labelled penicillin allergies compared to current practice. This program is likely to be self-sustainable from a funding perspective, as it is projected to result in significant future savings through reduced length of stay, optimised antibiotic utilisation, and reduced complication rates in future hospital admissions.
Statement of Ethics
This study protocol was reviewed and approved by the Central Adelaide Local Health Network Human Research Ethics Committee, Approval Nos. 15451 and 16325. Artificial intelligence screening was conducted without requiring written informed consent, since this component of the project was granted a waiver of consent by the Central Adelaide Local Health Network Human Research Ethics Committee. Written informed consent was required, and obtained, for in-person penicillin allergy delabelling evaluation.
Conflict of Interest Statement
The authors have no conflicts of interest to declare.
Funding Sources
This research is supported by a Central Adelaide Local Health Network Clinical Rapid Implementation Project Scheme (CRIPS) Grant (Reference No.: 17075). Stephen Bacchi is supported by a Fulbright Scholarship. No other funding was received for conducting this study or to assist with the preparation of this manuscript.
Author Contributions
M.J. conducted data collection and analysis and prepared the first draft. B.S., J.K., J.M.I., S.T., C.Y., S.S., and W.S. contributed to the conception of the idea for the work, study design, and critical manuscript review for important intellectual content. S.B. conducted data collection, analysis, and machine learning for the work.
Additional Information
Edited by: H.-U. Simon, Bern.
Data Availability Statement
Further enquiries can be made to the corresponding author. The data that support the findings of this study are not publicly available as they contain information that could compromise the privacy of research participants but are available in de-identified form from the corresponding author (M.J.) upon reasonable request.