Towards pharmacokinetic boosting of phenoxymethylpenicillin (penicillin-V) using probenecid for the treatment of bacterial infections.

In the face of increasing antimicrobial tolerance and resistance there is a global obligation to optimise oral antimicrobial dosing strategies including narrow spectrum penicillins, such as penicillin-V. We conducted a randomised, crossover study in healthy volunteers to characterise the influence of probenecid on penicillin-V pharmacokinetics and estimate the pharmacodynamics against Streptococcus pneumoniae. Twenty participants took six doses of penicillin-V (250 mg, 500 mg or 750 mg four times daily) with and without probenecid. Total and free concentrations of penicillin-V and probenecid were measured at two timepoints. A pharmacokinetic model was developed, and the probability of target attainment (PTA) calculated. The mean difference (95% CI) between penicillin-V alone and in combination with probenecid for serum total and free penicillin-V concentrations was significantly different at both timepoints (total: 45 min 4.32 (3.20-5.32) mg/L p < 0.001, 180 min 2.2 (1.58-3.25) mg/L p < 0.001; free: 45 min 1.15 (0.88-1.42) mg/L p < 0.001, 180 min 0.5 (0.35-0.76) mg/L p < 0.001). There was no difference between the timepoints in probenecid concentrations. PTA analysis shows probenecid allows a fourfold increase in MIC cover. Addition of probenecid was safe and well tolerated. The data support further research into improved dosing structures for complex outpatient therapy and might also be used to address penicillin supply shortages.

Towards a machine-learning assisted non-invasive classification of dengue severity using wearable PPG data: a prospective clinical study.

BACKGROUND: Dengue epidemics impose considerable strain on healthcare resources. Real-time continuous and non-invasive monitoring of patients admitted to the hospital could lead to improved care and outcomes. We evaluated the performance of a commercially available wearable (SmartCare) utilising photoplethysmography (PPG) to stratify clinical risk for a cohort of hospitalised patients with dengue in Vietnam. METHODS: We performed a prospective observational study for adult and paediatric patients with a clinical diagnosis of dengue at the Hospital for Tropical Disease, Ho Chi Minh City, Vietnam. Patients underwent PPG monitoring early during admission alongside standard clinical care. PPG waveforms were analysed using machine learning models. Adult patients were classified between 3 severity classes: i) uncomplicated (ward-based), ii) moderate-severe (emergency department-based), and iii) severe (ICU-based). Data from paediatric patients were split into 2 classes: i) severe (during ICU stay) and ii) follow-up (14-21 days after the illness onset). Model performances were evaluated using standard classification metrics and 5-fold stratified cross-validation. FINDINGS: We included PPG and clinical data from 132 adults and 15 paediatric patients with a median age of 28 (IQR, 21-35) and 12 (IQR, 9-13) years respectively. 1781 h of PPG data were available for analysis. The best performing convolutional neural network models (CNN) achieved a precision of 0.785 and recall of 0.771 in classifying adult patients according to severity class and a precision of 0.891 and recall of 0.891 in classifying between disease and post-disease state in paediatric patients. INTERPRETATION: We demonstrate that the use of a low-cost wearable provided clinically actionable data to differentiate between patients with dengue of varying severity. Continuous monitoring and connectivity to early warning systems could significantly benefit clinical care in dengue, particularly within an endemic setting. Work is currently underway to implement these models for dynamic risk predictions and assist in individualised patient care. FUNDING: EPSRC Centre for Doctoral Training in High-Performance Embedded and Distributed Systems (HiPEDS) (Grant: EP/L016796/1) and the Wellcome Trust (Grants: 215010/Z/18/Z and 215688/Z/19/Z).

Antimicrobial stewardship and targeted therapies in the changing landscape of maternal sepsis.

Pregnant and postnatal women are a high-risk population particularly prone to rapid progression to sepsis with significant morbidity and mortality worldwide. Moreover, severe maternal infections can have a serious detrimental impact on neonates with almost 1 million neonatal deaths annually attributed to maternal infection or sepsis. In this review we discuss the susceptibility of pregnant women and their specific physiological and immunological adaptations that contribute to their vulnerability to sepsis, the implications for the neonate, as well as the issues with antimicrobial stewardship and the challenges this poses when attempting to reach a balance between clinical care and urgent treatment. Finally, we review advancements in the development of pregnancy-specific diagnostic and therapeutic approaches and how these can be used to optimize the care of pregnant women and neonates.

Applying artificial intelligence and digital health technologies, Viet Nam.

Problem: Direct application of digital health technologies from high-income settings to low- and middle-income countries may be inappropriate due to challenges around data availability, implementation and regulation. Hence different approaches are needed. Approach: Within the Viet Nam ICU Translational Applications Laboratory project, since 2018 we have been developing a wearable device for individual patient monitoring and a clinical assessment tool to improve dengue disease management. Working closely with local staff at the Hospital for Tropical Diseases, Ho Chi Minh City, we developed and tested a prototype of the wearable device. We obtained perspectives on design and use of the sensor from patients. To develop the assessment tool, we used existing research data sets, mapped workflows and clinical priorities, interviewed stakeholders and held workshops with hospital staff. Local setting: In Viet Nam, a lower middle-income country, the health-care system is in the nascent stage of implementing digital health technologies. Relevant changes: Based on patient feedback, we are altering the design of the wearable sensor to increase comfort. We built the user interface of the assessment tool based on the core functionalities selected by workshop attendees. The interface was subsequently tested for usability in an iterative manner by the clinical staff members. Lessons learnt: The development and implementation of digital health technologies need an interoperable and appropriate plan for data management including collection, sharing and integration. Engagements and implementation studies should be conceptualized and conducted alongside the digital health technology development. The priorities of end-users, and understanding context and regulatory landscape are crucial for success.

A Multi-Site, Multi-Wavelength PPG Platform for Continuous Non-Invasive Health Monitoring in Hospital Settings.

This article presents a novel PPG acquisition platform capable of synchronous multi-wavelength signal acquisition from two measurement locations with up to 4 independent wavelengths from each in parallel. The platform is fully configurable and operates at 1ksps, accommodating a wide variety of transmitters and detectors to serve as both a research tool for experimentation and a clinical tool for disease monitoring. The sensing probes presented in this work acquire 4 PPG channels from the wrist and 4 PPG channels from the fingertip, with wavelengths such that surrogates for pulse wave velocity and haematocrit can be extracted. For conventional PPG sensing, we have achieved the mean error of 4.08 ± 3.72 bpm for heart-rate and a mean error of 1.54 ± 1.04% for SpO 2 measurement, with the latter lying within the FDA limits for commercial pulse oximeters. We have further evaluated over 700 individual peak-to-peak time differences between wrist and fingertip signals, achieving a normalized weighted average PWV of 5.80 ± 1.58 m/s, matching with values of PWV found for this age group in literature. Lastly, we introduced and computed a haematocrit ratio ( Rhct) between the deep IR and deep red wavelength from the fingertip sensor, finding a significant difference between male and female values (median of 1.9 and 2.93 respectively) pointing to devices sensitivity to Hct.

Learning meaningful latent space representations for patient risk stratification: Model development and validation for dengue and other acute febrile illness.

Background: Increased data availability has prompted the creation of clinical decision support systems. These systems utilise clinical information to enhance health care provision, both to predict the likelihood of specific clinical outcomes or evaluate the risk of further complications. However, their adoption remains low due to concerns regarding the quality of recommendations, and a lack of clarity on how results are best obtained and presented. Methods: We used autoencoders capable of reducing the dimensionality of complex datasets in order to produce a 2D representation denoted as latent space to support understanding of complex clinical data. In this output, meaningful representations of individual patient profiles are spatially mapped in an unsupervised manner according to their input clinical parameters. This technique was then applied to a large real-world clinical dataset of over 12,000 patients with an illness compatible with dengue infection in Ho Chi Minh City, Vietnam between 1999 and 2021. Dengue is a systemic viral disease which exerts significant health and economic burden worldwide, and up to 5% of hospitalised patients develop life-threatening complications. Results: The latent space produced by the selected autoencoder aligns with established clinical characteristics exhibited by patients with dengue infection, as well as features of disease progression. Similar clinical phenotypes are represented close to each other in the latent space and clustered according to outcomes broadly described by the World Health Organisation dengue guidelines. Balancing distance metrics and density metrics produced results covering most of the latent space, and improved visualisation whilst preserving utility, with similar patients grouped closer together. In this case, this balance is achieved by using the sigmoid activation function and one hidden layer with three neurons, in addition to the latent dimension layer, which produces the output (Pearson, 0.840; Spearman, 0.830; Procrustes, 0.301; GMM 0.321). Conclusion: This study demonstrates that when adequately configured, autoencoders can produce two-dimensional representations of a complex dataset that conserve the distance relationship between points. The output visualisation groups patients with clinically relevant features closely together and inherently supports user interpretability. Work is underway to incorporate these findings into an electronic clinical decision support system to guide individual patient management.

Continuous Measurement of Lactate Concentration in Human Subjects through Direct Electron Transfer from Enzymes to Microneedle Electrodes.

Microneedle lactate sensors may be used to continuously measure lactate concentration in the interstitial fluid in a minimally invasive and pain-free manner. First- and second-generation enzymatic sensors produce a redox-active product that is electrochemically sensed at the electrode surface. Direct electron transfer enzymes produce electrons directly as the product of enzymatic action; in this study, a direct electron transfer enzyme specific to lactate has been immobilized onto a microneedle surface to create lactate-sensing devices that function at low applied voltages (0.2 V). These devices have been validated in a small study of human volunteers; lactate concentrations were raised and lowered through physical exercise and subsequent rest. Lactazyme microneedle devices show good agreement with concurrently obtained and analyzed serum lactate levels.

Mapping patient pathways and understanding clinical decision-making in dengue management to inform the development of digital health tools.

BACKGROUND: Dengue is a common viral illness and severe disease results in life-threatening complications. Healthcare services in low- and middle-income countries treat the majority of dengue cases worldwide. However, the clinical decision-making processes which result in effective treatment are poorly characterised within this setting. In order to improve clinical care through interventions relating to digital clinical decision-support systems (CDSS), we set out to establish a framework for clinical decision-making in dengue management to inform implementation. METHODS: We utilised process mapping and task analysis methods to characterise existing dengue management at the Hospital for Tropical Diseases, Ho Chi Minh City, Vietnam. This is a tertiary referral hospital which manages approximately 30,000 patients with dengue each year, accepting referrals from Ho Chi Minh city and the surrounding catchment area. Initial findings were expanded through semi-structured interviews with clinicians in order to understand clinical reasoning and cognitive factors in detail. A grounded theory was used for coding and emergent themes were developed through iterative discussions with clinician-researchers. RESULTS: Key clinical decision-making points were identified: (i) at the initial patient evaluation for dengue diagnosis to decide on hospital admission and the provision of fluid/blood product therapy, (ii) in those patients who develop severe disease or other complications, (iii) at the point of recurrent shock in balancing the need for fluid therapy with complications of volume overload. From interviews the following themes were identified: prioritising clinical diagnosis and evaluation over existing diagnostics, the role of dengue guidelines published by the Ministry of Health, the impact of seasonality and caseload on decision-making strategies, and the potential role of digital decision-support and disease scoring tools. CONCLUSIONS: The study highlights the contemporary priorities in delivering clinical care to patients with dengue in an endemic setting. Key decision-making processes and the sources of information that were of the greatest utility were identified. These findings serve as a foundation for future clinical interventions and improvements in healthcare. Understanding the decision-making process in greater detail also allows for development and implementation of CDSS which are suited to the local context.

Rapid Measurement of Lactate in the Exhaled Breath Condensate: Biosensor Optimization and In-Human Proof of Concept.

Lactate concentration is of increasing interest as a diagnostic for sepsis, septic shock, and trauma. Compared with the traditional blood sample media, the exhaled breath condensate (EBC) has the advantages of non-invasiveness and higher user acceptance. An amperometric biosensor was developed and its application in EBC lactate detection was investigated in this paper. The sensor was modified with PEDOT:PSS-PB, and two different lactate oxidases (LODs). A rotating disk electrode and Koutecky-Levich analysis were applied for the kinetics analysis and gel optimization. The optimized gel formulation was then tested on disposable screen-printed sensors. The disposable sensors exhibited good performance and presented a high stability for both LOD modifications. Finally, human EBC analysis was conducted from a healthy subject at rest and after 30 min of intense aerobic cycling exercise. The sensor coulometric measurements showed good agreement with fluorometric and triple quadrupole liquid chromatography mass spectrometry reference methods. The EBC lactate concentration increased from 22.5 µM (at rest) to 28.0 µM (after 30 min of cycling) and dropped back to 5.3 µM after 60 min of rest.

A modified Sequential Organ Failure Assessment score for dengue: development, evaluation and proposal for use in clinical trials.

BACKGROUND: Dengue is a neglected tropical disease, for which no therapeutic agents have shown clinical efficacy to date. Clinical trials have used strikingly variable clinical endpoints, which hampers reproducibility and comparability of findings. We investigated a delta modified Sequential Organ Failure Assessment (delta mSOFA) score as a uniform composite clinical endpoint for use in clinical trials investigating therapeutics for moderate and severe dengue. METHODS: We developed a modified SOFA score for dengue, measured and evaluated its performance at baseline and 48 h after enrolment in a prospective observational cohort of 124 adults admitted to a tertiary referral hospital in Vietnam with dengue shock. The modified SOFA score included pulse pressure in the cardiovascular component. Binary logistic regression, cox proportional hazard and linear regression models were used to estimate association between mSOFA, delta mSOFA and clinical outcomes. RESULTS: The analysis included 124 adults with dengue shock. 29 (23.4%) patients required ICU admission for organ support or due to persistent haemodynamic instability: 9/124 (7.3%) required mechanical ventilation, 8/124 (6.5%) required vasopressors, 6/124 (4.8%) required haemofiltration and 5/124 (4.0%) patients died. In univariate analyses, higher baseline and delta (48 h) mSOFA score for dengue were associated with admission to ICU, requirement for organ support and mortality, duration of ICU and hospital admission and IV fluid use. CONCLUSIONS: The baseline and delta mSOFA scores for dengue performed well to discriminate patients with dengue shock by clinical outcomes, including duration of ICU and hospital admission, requirement for organ support and death. We plan to use delta mSOFA as the primary endpoint in an upcoming host-directed therapeutic trial and investigate the performance of this score in other phenotypes of severe dengue in adults and children.

Prediction of hospital-onset COVID-19 infections using dynamic networks of patient contact: an international retrospective cohort study.

BACKGROUND: Real-time prediction is key to prevention and control of infections associated with health-care settings. Contacts enable spread of many infections, yet most risk prediction frameworks fail to account for their dynamics. We developed, tested, and internationally validated a real-time machine-learning framework, incorporating dynamic patient-contact networks to predict hospital-onset COVID-19 infections (HOCIs) at the individual level. METHODS: We report an international retrospective cohort study of our framework, which extracted patient-contact networks from routine hospital data and combined network-derived variables with clinical and contextual information to predict individual infection risk. We trained and tested the framework on HOCIs using the data from 51 157 hospital inpatients admitted to a UK National Health Service hospital group (Imperial College Healthcare NHS Trust) between April 1, 2020, and April 1, 2021, intersecting the first two COVID-19 surges. We validated the framework using data from a Swiss hospital group (Department of Rehabilitation, Geneva University Hospitals) during a COVID-19 surge (from March 1 to May 31, 2020; 40 057 inpatients) and from the same UK group after COVID-19 surges (from April 2 to Aug 13, 2021; 43 375 inpatients). All inpatients with a bed allocation during the study periods were included in the computation of network-derived and contextual variables. In predicting patient-level HOCI risk, only inpatients spending 3 or more days in hospital during the study period were examined for HOCI acquisition risk. FINDINGS: The framework was highly predictive across test data with all variable types (area under the curve [AUC]-receiver operating characteristic curve [ROC] 0·89 [95% CI 0·88-0·90]) and similarly predictive using only contact-network variables (0·88 [0·86-0·90]). Prediction was reduced when using only hospital contextual (AUC-ROC 0·82 [95% CI 0·80-0·84]) or patient clinical (0·64 [0·62-0·66]) variables. A model with only three variables (ie, network closeness, direct contacts with infectious patients [network derived], and hospital COVID-19 prevalence [hospital contextual]) achieved AUC-ROC 0·85 (95% CI 0·82-0·88). Incorporating contact-network variables improved performance across both validation datasets (AUC-ROC in the Geneva dataset increased from 0·84 [95% CI 0·82-0·86] to 0·88 [0·86-0·90]; AUC-ROC in the UK post-surge dataset increased from 0·49 [0·46-0·52] to 0·68 [0·64-0·70]). INTERPRETATION: Dynamic contact networks are robust predictors of individual patient risk of HOCIs. Their integration in clinical care could enhance individualised infection prevention and early diagnosis of COVID-19 and other nosocomial infections. FUNDING: Medical Research Foundation, WHO, Engineering and Physical Sciences Research Council, National Institute for Health Research (NIHR), Swiss National Science Foundation, and German Research Foundation.

The Diagnosis of Dengue in Patients Presenting With Acute Febrile Illness Using Supervised Machine Learning and Impact of Seasonality.

Background: Symptomatic dengue infection can result in a life-threatening shock syndrome and timely diagnosis is essential. Point-of-care tests for non-structural protein 1 and IgM are used widely but performance can be limited. We developed a supervised machine learning model to predict whether patients with acute febrile illnesses had a diagnosis of dengue or other febrile illnesses (OFI). The impact of seasonality on model performance over time was examined. Methods: We analysed data from a prospective observational clinical study in Vietnam. Enrolled patients presented with an acute febrile illness of <72 h duration. A gradient boosting model (XGBoost) was used to predict final diagnosis using age, sex, haematocrit, platelet, white cell, and lymphocyte count collected on enrolment. Data was randomly split 80/20% into a training and hold-out set, respectively, with the latter not used in model development. Cross-validation and hold out set testing was used, with performance over time evaluated through a rolling window approach. Results: We included 8,100 patients recruited between 16th October 2010 and 10th December 2014. In total 2,240 (27.7%) patients were diagnosed with dengue infection. The optimised model from training data had an overall median area under the receiver operator curve (AUROC) of 0.86 (interquartile range 0.84-0.86), specificity of 0.92, sensitivity of 0.56, positive predictive value of 0.73, negative predictive value (NPV) of 0.84, and Brier score of 0.13 in predicting the final diagnosis, with similar performances in hold-out set testing (AUROC of 0.86). Model performances varied significantly over time as a function of seasonality and other factors. Incorporation of a dynamic threshold which continuously learns from recent cases resulted in a more consistent performance throughout the year (NPV >90%). Conclusion: Supervised machine learning models are able to discriminate between dengue and OFI diagnoses in patients presenting with an early undifferentiated febrile illness. These models could be of clinical utility in supporting healthcare decision-making and provide passive surveillance across dengue endemic regions. Effects of seasonality and changing disease prevalence must however be taken into account-this is of significant importance given unpredictable effects of human-induced climate change and the impact on health.

The compensatory reserve index predicts recurrent shock in patients with severe dengue.

BACKGROUND: Dengue shock syndrome (DSS) is one of the major clinical phenotypes of severe dengue. It is defined by significant plasma leak, leading to intravascular volume depletion and eventually cardiovascular collapse. The compensatory reserve Index (CRI) is a new physiological parameter, derived from feature analysis of the pulse arterial waveform that tracks real-time changes in central volume. We investigated the utility of CRI to predict recurrent shock in severe dengue patients admitted to the ICU. METHODS: We performed a prospective observational study in the pediatric and adult intensive care units at the Hospital for Tropical Diseases, Ho Chi Minh City, Vietnam. Patients were monitored with hourly clinical parameters and vital signs, in addition to continuous recording of the arterial waveform using pulse oximetry. The waveform data was wirelessly transmitted to a laptop where it was synchronized with the patient's clinical data. RESULTS: One hundred three patients with suspected severe dengue were recruited to this study. Sixty-three patients had the minimum required dataset for analysis. Median age was 11 years (IQR 8-14 years). CRI had a negative correlation with heart rate and moderate negative association with blood pressure. CRI was found to predict recurrent shock within 12 h of being measured (OR 2.24, 95% CI 1.54-3.26), P < 0.001). The median duration from CRI measurement to the first recurrent shock was 5.4 h (IQR 2.9-6.8). A CRI cutoff of 0.4 provided the best combination of sensitivity and specificity for predicting recurrent shock (0.66 [95% CI 0.47-0.85] and 0.86 [95% CI 0.80-0.92] respectively). CONCLUSION: CRI is a useful non-invasive method for monitoring intravascular volume status in patients with severe dengue.

Applied machine learning for the risk-stratification and clinical decision support of hospitalised patients with dengue in Vietnam.

BACKGROUND: Identifying patients at risk of dengue shock syndrome (DSS) is vital for effective healthcare delivery. This can be challenging in endemic settings because of high caseloads and limited resources. Machine learning models trained using clinical data could support decision-making in this context. METHODS: We developed supervised machine learning prediction models using pooled data from adult and paediatric patients hospitalised with dengue. Individuals from 5 prospective clinical studies in Ho Chi Minh City, Vietnam conducted between 12th April 2001 and 30th January 2018 were included. The outcome was onset of dengue shock syndrome during hospitalisation. Data underwent random stratified splitting at 80:20 ratio with the former used only for model development. Ten-fold cross-validation was used for hyperparameter optimisation and confidence intervals derived from percentile bootstrapping. Optimised models were evaluated against the hold-out set. FINDINGS: The final dataset included 4,131 patients (477 adults and 3,654 children). DSS was experienced by 222 (5.4%) of individuals. Predictors were age, sex, weight, day of illness at hospitalisation, indices of haematocrit and platelets over first 48 hours of admission and before the onset of DSS. An artificial neural network model (ANN) model had best performance with an area under receiver operator curve (AUROC) of 0.83 (95% confidence interval [CI], 0.76-0.85) in predicting DSS. When evaluated against the independent hold-out set this calibrated model exhibited an AUROC of 0.82, specificity of 0.84, sensitivity of 0.66, positive predictive value of 0.18 and negative predictive value of 0.98. INTERPRETATION: The study demonstrates additional insights can be obtained from basic healthcare data, when applied through a machine learning framework. The high negative predictive value could support interventions such as early discharge or ambulatory patient management in this population. Work is underway to incorporate these findings into an electronic clinical decision support system to guide individual patient management.

Informing antimicrobial management in the context of COVID-19: understanding the longitudinal dynamics of C-reactive protein and procalcitonin.

BACKGROUND: To characterise the longitudinal dynamics of C-reactive protein (CRP) and Procalcitonin (PCT) in a cohort of hospitalised patients with COVID-19 and support antimicrobial decision-making. METHODS: Longitudinal CRP and PCT concentrations and trajectories of 237 hospitalised patients with COVID-19 were modelled. The dataset comprised of 2,021 data points for CRP and 284 points for PCT. Pairwise comparisons were performed between: (i) those with or without significant bacterial growth from cultures, and (ii) those who survived or died in hospital. RESULTS: CRP concentrations were higher over time in COVID-19 patients with positive microbiology (day 9: 236 vs 123 mg/L, p < 0.0001) and in those who died (day 8: 226 vs 152 mg/L, p < 0.0001) but only after day 7 of COVID-related symptom onset. Failure for CRP to reduce in the first week of hospital admission was associated with significantly higher odds of death. PCT concentrations were higher in patients with COVID-19 and positive microbiology or in those who died, although these differences were not statistically significant. CONCLUSIONS: Both the absolute CRP concentration and the trajectory during the first week of hospital admission are important factors predicting microbiology culture positivity and outcome in patients hospitalised with COVID-19. Further work is needed to describe the role of PCT for co-infection. Understanding relationships of these biomarkers can support development of risk models and inform optimal antimicrobial strategies.

Combination of inflammatory and vascular markers in the febrile phase of dengue is associated with more severe outcomes.

Background: Early identification of severe dengue patients is important regarding patient management and resource allocation. We investigated the association of 10 biomarkers (VCAM-1, SDC-1, Ang-2, IL-8, IP-10, IL-1RA, sCD163, sTREM-1, ferritin, CRP) with the development of severe/moderate dengue (S/MD). Methods: We performed a nested case-control study from a multi-country study. A total of 281 S/MD and 556 uncomplicated dengue cases were included. Results: On days 1-3 from symptom onset, higher levels of any biomarker increased the risk of developing S/MD. When assessing together, SDC-1 and IL-1RA were stable, while IP-10 changed the association from positive to negative; others showed weaker associations. The best combinations associated with S/MD comprised IL-1RA, Ang-2, IL-8, ferritin, IP-10, and SDC-1 for children, and SDC-1, IL-8, ferritin, sTREM-1, IL-1RA, IP-10, and sCD163 for adults. Conclusions: Our findings assist the development of biomarker panels for clinical use and could improve triage and risk prediction in dengue patients. Funding: This study was supported by the EU's Seventh Framework Programme (FP7-281803 IDAMS), the WHO, and the Bill and Melinda Gates Foundation.

Supervised machine learning to support the diagnosis of bacterial infection in the context of COVID-19.

BACKGROUND: Bacterial infection has been challenging to diagnose in patients with COVID-19. We developed and evaluated supervised machine learning algorithms to support the diagnosis of secondary bacterial infection in hospitalized patients during the COVID-19 pandemic. METHODS: Inpatient data at three London hospitals for the first COVD-19 wave in March and April 2020 were extracted. Demographic, blood test and microbiology data for individuals with and without SARS-CoV-2-positive PCR were obtained. A Gaussian Naive Bayes, Support Vector Machine (SVM) and Artificial Neural Network were trained and compared using the area under the receiver operating characteristic curve (AUCROC). The best performing algorithm (SVM with 21 blood test variables) was prospectively piloted in July 2020. AUCROC was calculated for the prediction of a positive microbiological sample within 48 h of admission. RESULTS: A total of 15 599 daily blood profiles for 1186 individual patients were identified to train the algorithms; 771/1186 (65%) individuals were SARS-CoV-2 PCR positive. Clinically significant microbiology results were present for 166/1186 (14%) patients during admission. An SVM algorithm trained with 21 routine blood test variables and over 8000 individual profiles had the best performance. AUCROC was 0.913, sensitivity 0.801 and specificity 0.890. Prospective testing on 54 patients on admission (28/54, 52% SARS-CoV-2 PCR positive) demonstrated an AUCROC of 0.960 (95% CI: 0.90-1.00). CONCLUSIONS: An SVM using 21 routine blood test variables had excellent performance at inferring the likelihood of positive microbiology. Further prospective evaluation of the algorithms ability to support decision making for the diagnosis of bacterial infection in COVID-19 cohorts is underway.