Explainable and Interpretable Machine Learning for Antimicrobial Stewardship: Opportunities and Challenges.

There is growing interest in exploiting the advances in artificial intelligence and machine learning (ML) for improving and monitoring antimicrobial prescriptions in line with antimicrobial stewardship principles. Against this background, the concepts of interpretability and explainability are becoming increasingly essential to understanding how ML algorithms could predict antimicrobial resistance or recommend specific therapeutic agents, to avoid unintended biases related to the "black box" nature of complex models. In this commentary, we review and discuss some relevant topics on the use of ML algorithms for antimicrobial stewardship interventions, highlighting opportunities and challenges, with particular attention paid to interpretability and explainability of employed models. As in other fields of medicine, the exponential growth of artificial intelligence and ML indicates the potential for improving the efficacy of antimicrobial stewardship interventions, at least in part by reducing time-consuming tasks for overwhelmed health care personnel. Improving our knowledge about how complex ML models work could help to achieve crucial advances in promoting the appropriate use of antimicrobials, as well as in preventing antimicrobial resistance selection and dissemination.

Early diagnosis of candidemia with explainable machine learning on automatically extracted laboratory and microbiological data: results of the AUTO-CAND project.

BACKGROUND: Candidemia is associated with a heavy burden of morbidity and mortality in hospitalized patients. The availability of blood culture results could require up to 48-72 h after blood draw; thus, early treatment decisions are made in the absence of a definite diagnosis. METHODS: In this retrospective study, we assessed the performance of different supervised machine learning algorithms for the early differential diagnosis of candidemia and bacteremia in adult patients on a large dataset automatically extracted within the AUTO-CAND project. RESULTS: Overall, 12,483 episodes of candidemia (1275; 10%) or bacteremia (11,208; 90%) were included in the analysis. A random forest classifier achieved the best diagnostic performance for candidemia, with sensitivity 0.98 and specificity 0.65 on the training set (true skill statistic [TSS] = 0.63) and sensitivity 0.74 and specificity 0.57 on the test set (TSS = 0.31). Then, the random classifier was trained in the subgroup of patients with available serum ß-D-glucan (BDG) and procalcitonin (PCT) values by exploiting the feature ranking learned in the entire dataset. Although no statistically significant differences were observed from the performance measures obtained by employing BDG and PCT alone, the performance measures of the classifier that included the features selected in the entire dataset, plus BDG and PCT, were the highest in most cases. CONCLUSIONS: Random forest classifiers trained on large datasets of automatically extracted data have the potential to improve current diagnostic algorithms for candidemia. However, further development through implementation of automatically extracted clinical features may be necessary to achieve crucial improvements.

Validation of an Automated System for the Extraction of a Wide Dataset for Clinical Studies Aimed at Improving the Early Diagnosis of Candidemia.

There is increasing interest in assessing whether machine learning (ML) techniques could further improve the early diagnosis of candidemia among patients with a consistent clinical picture. The objective of the present study is to validate the accuracy of a system for the automated extraction from a hospital laboratory software of a large number of features from candidemia and/or bacteremia episodes as the first phase of the AUTO-CAND project. The manual validation was performed on a representative and randomly extracted subset of episodes of candidemia and/or bacteremia. The manual validation of the random extraction of 381 episodes of candidemia and/or bacteremia, with automated organization in structured features of laboratory and microbiological data resulted in ≥99% correct extractions (with confidence interval < ±1%) for all variables. The final automatically extracted dataset consisted of 1338 episodes of candidemia (8%), 14,112 episodes of bacteremia (90%), and 302 episodes of mixed candidemia/bacteremia (2%). The final dataset will serve to assess the performance of different ML models for the early diagnosis of candidemia in the second phase of the AUTO-CAND project.

Prediction of severe thunderstorm events with ensemble deep learning and radar data.

The problem of nowcasting extreme weather events can be addressed by applying either numerical methods for the solution of dynamic model equations or data-driven artificial intelligence algorithms. Within this latter framework, the most used techniques rely on video prediction deep learning methods which take in input time series of radar reflectivity images to predict the next future sequence of reflectivity images, from which the predicted rainfall quantities are extrapolated. Differently from the previous works, the present paper proposes a deep learning method, exploiting videos of radar reflectivity frames as input and lightning data to realize a warning machine able to sound timely alarms of possible severe thunderstorm events. The problem is recast in a classification one in which the extreme events to be predicted are characterized by a an high level of precipitation and lightning density. From a technical viewpoint, the computational core of this approach is an ensemble learning method based on the recently introduced value-weighted skill scores for both transforming the probabilistic outcomes of the neural network into binary predictions and assessing the forecasting performance. Such value-weighted skill scores are particularly suitable for binary predictions performed over time since they take into account the time evolution of events and predictions paying attention to the value of the prediction for the forecaster. The result of this study is a warning machine validated against weather radar data recorded in the Liguria region, in Italy.

Bad and Good Errors: Value-Weighted Skill Scores in Deep Ensemble Learning.

Forecast verification is a crucial task for assessing the predictive power of prognostic model forecasts and it is usually implemented by checking quality-based skill scores. In this article, we propose a novel approach to realize forecast verification focusing not just on the forecast quality but rather on its value. Specifically, we introduce a strategy for assessing the severity of forecast errors based on the evidence that, on the one hand, a false alarm just anticipating an occurring event is better than one in the middle of consecutive nonoccurring events, and that, on the other hand, a miss of an isolated event has a worse impact than a miss of a single event, which is part of several consecutive occurrences. Relying on this idea, we introduce a notion of value-weighted skill scores giving greater importance to the value of the prediction rather than to its quality. Then, we introduce an ensemble strategy to maximize quality-based and value-weighted skill scores independently of one another. We test it on the predictions provided by deep learning methods for binary classification in the case of four applications concerned with pollution, space weather, stock price, and IoT data stream forecasting. Our experimental studies show that using the ensemble strategy for maximizing the value-weighted skill scores generally improves both the value and quality of the forecast.