Pharmacometrics and Machine Learning Partner to Advance Clinical Data Analysis

Prediction of teicoplanin plasma concentration in critically ill patients: a combination of machine learning and population pharmacokinetics

BACKGROUND

Teicoplanin has been widely used in patients with infections caused by Staphylococcus aureus, especially for critically ill patients. The pharmacokinetics (PK) of teicoplanin vary between individuals and within the same individual. We aim to establish a prediction model via a combination of machine learning and population PK (PPK) to support personalized medication decisions for critically ill patients.

METHODS

A retrospective study was performed incorporating 33 variables, including PPK parameters (clearance and volume of distribution). Multiple algorithms and Shapley additive explanations were employed for feature selection of variables to determine the strongest driving factors.

RESULTS

The performance of each algorithm with PPK parameters was superior to that without PPK parameters. The composition of support vector regression, categorical boosting and a backpropagation neural network (7:2:1) with the highest R2 (0.809) was determined as the final ensemble model. The model included 15 variables after feature selection, of which the predictive performance was superior to that of models considering all variables or using only PPK. The R2, mean absolute error, mean squared error, absolute accuracy (±5 mg/L) and relative accuracy (±30%) of external validation were 0.649, 3.913, 28.347, 76.12% and 76.12%, respectively.

CONCLUSIONS

Our study offers a non-invasive, fast and cost-effective prediction model of teicoplanin plasma concentration in critically ill patients. The model serves as a fundamental tool for clinicians to determine the effective plasma concentration range of teicoplanin and formulate individualized dosing regimens accordingly.

pyDarwin machine learning algorithms application and comparison in nonlinear mixed-effect model selection and optimization

Forward addition/backward elimination (FABE) has been the standard for population pharmacokinetic model selection (PPK) since NONMEM® was introduced. We investigated five machine learning (ML) algorithms (Genetic algorithm [GA], Gaussian process [GP], random forest [RF], gradient boosted random tree [GBRT], and particle swarm optimization [PSO]) as alternatives to FABE. These algorithms were applied to PPK model selection with a focus on comparing the efficiency and robustness of each of them. All machine learning algorithms included the combination of ML algorithms with a local downhill search. The local downhill search consisted of systematically changing one or two "features" at a time (a one-bit or a two-bit local search), alternating with the ML methods. An exhaustive search (all possible combinations of model features, N = 1,572,864 models) was the gold standard for robustness, and the number of models examined leading prior to identification of the final model was the metric for efficiency.All algorithms identified the optimal model when combined with the two-bit local downhill search. GA, RF, GBRT, and GP identified the optimal model with only a one-bit local search. PSO required the two-bit local downhill search. In our analysis, GP was the most efficient algorithm as measured by the number of models examined prior to finding the optimal (495 models), and PSO exhibited the least efficiency, requiring 1710 unique models before finding the best solution. Additionally, GP was also the algorithm that needed the longest elapsed time of 2975.6 min, in comparison with GA, which only required 321.8 min.

Clustering plasma concentration-time curves: applications of unsupervised learning in pharmacogenomics

Pharmaceutical researchers are continually searching for techniques to improve both drug development processes and patient outcomes. An area of recent interest is the potential for machine learning (ML) applications within pharmacology. One such application not yet given close study is the unsupervised clustering of plasma concentration-time curves, hereafter, pharmacokinetic (PK) curves. In this paper, we present our findings on how to cluster PK curves by their similarity. Specifically, we find clustering to be effective at identifying similar-shaped PK curves and informative for understanding patterns within each cluster of PK curves. Because PK curves are time series data objects, our approach utilizes the extensive body of research related to the clustering of time series data as a starting point. As such, we examine many dissimilarity measures between time series data objects to find those most suitable for PK curves. We identify Euclidean distance as generally most appropriate for clustering PK curves, and we further show that dynamic time warping, Fréchet, and structure-based measures of dissimilarity like correlation may produce unexpected results. As an illustration, we apply these methods in a case study with 250 PK curves used in a previous pharmacogenomic study. Our case study finds that an unsupervised ML clustering with Euclidean distance, without any subject genetic information, is able to independently validate the same conclusions as the reference pharmacogenomic results. To our knowledge, this is the first such demonstration. Further, the case study demonstrates how the clustering of PK curves may generate insights that could be difficult to perceive solely with population level summary statistics of PK metrics.

pyDarwin: A Machine Learning Enhanced Automated Nonlinear Mixed-Effect Model Selection Toolbox

pyDarwin is an open-source Python package for nonlinear mixed-effect model selection. pyDarwin combines machine-learning algorithms and NONMEM to perform a global search for the optimal model in a user-defined model search space. Compared with traditional stepwise search, pyDarwin provides an efficient platform for conducting an objective, robust, less labor-intensive model selection process without compromising model interpretability. In this tutorial, we will begin by introducing the essential components and concepts within the package. Subsequently, we will provide an overview of the pyDarwin modeling workflow and the necessary files needed for model selection. To illustrate the entire process, we will conclude with an example utilizing quetiapine clinical data.

Low-dimensional neural ODEs and their application in pharmacokinetics

Machine Learning (ML) is a fast-evolving field, integrated in many of today's scientific disciplines. With the recent development of neural ordinary differential equations (NODEs), ML provides a new tool to model dynamical systems in the field of pharmacology and pharmacometrics, such as pharmacokinetics (PK) or pharmacodynamics. The novel and conceptionally different approach of NODEs compared to classical PK modeling creates challenges but also provides opportunities for its application. In this manuscript, we introduce the functionality of NODEs and develop specific low-dimensional NODE structures based on PK principles. We discuss two challenges of NODEs, overfitting and extrapolation to unseen data, and provide practical solutions to these problems. We illustrate concept and application of our proposed low-dimensional NODE approach with several PK modeling examples, including multi-compartmental, target-mediated drug disposition, and delayed absorption behavior. In all investigated scenarios, the NODEs were able to describe the data well and simulate data for new subjects within the observed dosing range. Finally, we briefly demonstrate how NODEs can be combined with mechanistic models. This research work enhances understanding of how NODEs can be applied in PK analyses and illustrates the potential for NODEs in the field of pharmacology and pharmacometrics.

Go beyond the limits of genetic algorithm in daily covariate selection practice

Covariate identification is an important step in the development of a population pharmacokinetic/pharmacodynamic model. Among the different available approaches, the stepwise covariate model (SCM) is the most used. However, SCM is based on a local search strategy, in which the model-building process iteratively tests the addition or elimination of a single covariate at a time given all the others. This introduces a heuristic to limit the searching space and then the computational complexity, but, at the same time, can lead to a suboptimal solution. The application of genetic algorithms (GAs) for covariate selection has been proposed as a possible solution to overcome these limitations. However, their actual use during model building is limited by the extremely high computational costs and convergence issues, both related to the number of models being tested. In this paper, we proposed a new GA for covariate selection to address these challenges. The GA was first developed on a simulated case study where the heuristics introduced to overcome the limitations affecting currently available GA approaches resulted able to limit the selection of redundant covariates, increase replicability of results and reduce convergence times. Then, we tested the proposed GA on a real-world problem related to remifentanil. It obtained good results both in terms of selected covariates and fitness optimization, outperforming the SCM.

Application of machine learning combined with population pharmacokinetics to improve individual prediction of vancomycin clearance in simulated adult patients

Background and aim

Vancomycin, a glycopeptide antimicrobial drug. PPK has problems such as difficulty in accurately reflecting inter-individual differences, and the PPK model may not be accurate enough to predict individual pharmacokinetic parameters. Therefore, the aim of this study is to investigate whether the application of machine learning combined with the PPK method can improve the prediction of vancomycin CL in adult Chinese patients.

Methods

In the first step, a vancomycin CL prediction model for Chinese adult patients is given by PPK and Hamilton Monte Carlo sampling is used to obtain the reference CL of 1,000 patients; the second step is to obtain the final prediction model by machine learning using an appropriate model for the predictive factor and the reference CL; and the third step is to randomly select, in the simulated data, a total of 250 patients for prediction effect evaluation.

Results

XGBoost model is selected as final machine learning model. More than four-fifths of the subjects' predictive values regarding vancomycin CL are improved by machine learning combined with PPK. Machine learning combined with PPK models is more stable in performance than the PPK method alone for predicting models.

Conclusion

The first combination of PPK and machine learning for predictive modeling of vancomycin clearance in adult patients. It provides a reference for clinical pharmacists or clinicians to optimize the initial dosage given to ensure the effectiveness and safety of drug therapy for each patient.

Investigating the value of glucodensity analysis of continuous glucose monitoring data in type 1 diabetes: an exploratory analysis

Introduction

Continuous glucose monitoring (CGM) devices capture longitudinal data on interstitial glucose levels and are increasingly used to show the dynamics of diabetes metabolism. Given the complexity of CGM data, it is crucial to extract important patterns hidden in these data through efficient visualization and statistical analysis techniques.

Methods

In this paper, we adopted the concept of glucodensity, and using a subset of data from an ongoing clinical trial in pediatric individuals and young adults with new-onset type 1 diabetes, we performed a cluster analysis of glucodensities. We assessed the differences among the identified clusters using analysis of variance (ANOVA) with respect to residual pancreatic beta-cell function and some standard CGM-derived parameters such as time in range, time above range, and time below range.

Results

Distinct CGM data patterns were identified using cluster analysis based on glucodensities. Statistically significant differences were shown among the clusters with respect to baseline levels of pancreatic beta-cell function surrogate (C-peptide) and with respect to time in range and time above range.

Discussion

Our findings provide supportive evidence for the value of glucodensity in the analysis of CGM data. Some challenges in the modeling of CGM data include unbalanced data structure, missing observations, and many known and unknown confounders, which speaks to the importance of--and provides opportunities for--taking an approach integrating clinical, statistical, and data science expertise in the analysis of these data.

Development of machine learning algorithms to estimate maximum residue limits for veterinary medicines

Establishing maximum-residue limits (MRLs) for veterinary medicine helps to protect the human food supply. Guidelines for establishing MRLs are outlined by regulatory authorities that drug sponsors follow in each country. During the drug approval process, residue limits are targeted for specific animal species and matrices. Therefore, MRLs are commonly not established for other species. This study demonstrates unestablished MRLs can be reliably predicted for under-represented food commodity groups using machine learning (ML). Classification methods with imbalanced data were used to analyze MRL data from multiple countries by implementing resampling techniques in different ML classifiers. Afterward, we developed and evaluated a data-mining method for predicting unestablished MRLs. Seven different ML classifiers such as support vector classifier, multi-layer perceptron (MLP), random forest, decision tree, k-neighbors, Gaussian NB, and AdaBoost have been selected in this baseline study. Among these, the neural network MLP classifier reliably scored the highest average-weighted F1 score (accuracy >99% with markers and ≈88% without markets) in predicting unestablished MRLs. This provides the first study to apply ML algorithms in regulatory food animal medicine. By predicting and estimating MRLs, we can potentially decrease the use and cost of live animals and the overall research burden of determining new MRLs.

Modeling tumor size dynamics based on real-world electronic health records and image data in advanced melanoma patients receiving immunotherapy

The development of immune checkpoint inhibitors (ICIs) has revolutionized cancer therapy but only a fraction of patients benefits from this therapy. Model-informed drug development can be used to assess prognostic and predictive clinical factors or biomarkers associated with treatment response. Most pharmacometric models have thus far been developed using data from randomized clinical trials, and further studies are needed to translate their findings into the real-world setting. We developed a tumor growth inhibition model based on real-world clinical and imaging data in a population of 91 advanced melanoma patients receiving ICIs (i.e., ipilimumab, nivolumab, and pembrolizumab). Drug effect was modeled as an ON/OFF treatment effect, with a tumor killing rate constant identical for the three drugs. Significant and clinically relevant covariate effects of albumin, neutrophil to lymphocyte ratio, and Eastern Cooperative Oncology Group (ECOG) performance status were identified on the baseline tumor volume parameter, as well as NRAS mutation on tumor growth rate constant using standard pharmacometric approaches. In a population subgroup (n = 38), we had the opportunity to conduct an exploratory analysis of image-based covariates (i.e., radiomics features), by combining machine learning and conventional pharmacometric covariate selection approaches. Overall, we demonstrated an innovative pipeline for longitudinal analyses of clinical and imaging RWD with a high-dimensional covariate selection method that enabled the identification of factors associated with tumor dynamics. This study also provides a proof of concept for using radiomics features as model covariates.

Virtual patients, digital twins and causal disease models: paving the ground for in silico clinical trials

Computational models are being explored to simulate in silico the efficacy and safety of drug candidates and medical devices. Disease models that are based on patients' profiling data are being produced to represent interactomes of genes or proteins and to infer causality in the pathophysiology {AuQ: Edit OK?}, which makes it possible to mimic the impact of drugs on relevant targets. Virtual patients designed from medical records as well as digital twins were generated to simulate specific organs and to predict treatment efficacy at the individual patient level {AuQ: Edit OK?}. As the acceptance of digital evidence by regulators grows, predictive artificial intelligence (AI)-based models will support the design of confirmatory trials in humans and will accelerate the development of efficient drugs and medical devices.

Deep Learning Methods Applied to Drug Concentration Prediction of Olanzapine

Pharmacometrics and the utilization of population pharmacokinetics play an integral role in model-informed drug discovery and development (MIDD). Recently, there has been a growth in the application of deep learning approaches to aid in areas within MIDD. In this study, a deep learning model, LSTM-ANN, was developed to predict olanzapine drug concentrations from the CATIE study. A total of 1527 olanzapine drug concentrations from 523 individuals along with 11 patient-specific covariates were used in model development. The hyperparameters of the LSTM-ANN model were optimized through a Bayesian optimization algorithm. A population pharmacokinetic model using the NONMEM model was constructed as a reference to compare to the performance of the LSTM-ANN model. The RMSE of the LSTM-ANN model was 29.566 in the validation set, while the RMSE of the NONMEM model was 31.129. Permutation importance revealed that age, sex, and smoking were highly influential covariates in the LSTM-ANN model. The LSTM-ANN model showed potential in the application of drug concentration predictions as it was able to capture the relationships within a sparsely sampled pharmacokinetic dataset and perform comparably to the NONMEM model.

A Hybrid Algorithm Combining Population Pharmacokinetic and Machine Learning for Isavuconazole Exposure Prediction

OBJECTIVES

Maximum a posteriori Bayesian estimation (MAP-BE) based on a limited sampling strategy and a population pharmacokinetic (POPPK) model is used to estimate individual pharmacokinetic parameters. Recently, we proposed a methodology that combined population pharmacokinetic and machine learning (ML) to decrease the bias and imprecision in individual iohexol clearance prediction. The aim of this study was to confirm the previous results by developing a hybrid algorithm combining POPPK, MAP-BE and ML that accurately predicts isavuconazole clearance.

METHODS

A total of 1727 isavuconazole rich PK profiles were simulated using a POPPK model from the literature, and MAP-BE was used to estimate the clearance based on: (i) the full PK profiles (refCL); and (ii) C24h only (C24h-CL). Xgboost was trained to correct the error between refCL and C24h-CL in the training dataset (75%). C24h-CL as well as ML-corrected C24h-CL were evaluated in a testing dataset (25%) and then in a set of PK profiles simulated using another published POPPK model.

RESULTS

A strong decrease in mean predictive error (MPE%), imprecision (RMSE%) and the number of profiles outside ± 20% MPE% (n-out20%) was observed with the hybrid algorithm (decreased in MPE% by 95.8% and 85.6%; RMSE% by 69.5% and 69.0%; n-out20% by 97.4% and 100% in the training and testing sets, respectively. In the external validation set, the hybrid algorithm decreased MPE% by 96%, RMSE% by 68% and n-out20% by 100%.

CONCLUSION

The hybrid model proposed significantly improved isavuconazole AUC estimation over MAP-BE based on the sole C24h and may improve dose adjustment.

Applying machine learning to the pharmacokinetic modeling of cyclosporine in adult renal transplant recipients: a multi-method comparison

Objective: The aim of this study was to identify the important factors affecting cyclosporine (CsA) blood concentration and estimate CsA concentration using seven different machine learning (ML) algorithms. We also assessed the predictability of established ML models and previously built population pharmacokinetic (popPK) model. Finally, the most suitable ML model and popPK model to guide precision dosing were determined.

Methods: In total, 3,407 whole-blood trough and peak concentrations of CsA were obtained from 183 patients who underwent initial renal transplantation. These samples were divided into model-building and evaluation sets. The model-building set was analyzed using seven different ML algorithms. The effects of potential covariates were evaluated using the least absolute shrinkage and selection operator algorithms. A separate evaluation set was used to assess the ability of all models to predict CsA blood concentration. R squared (R2) scores, median prediction error (MDPE), median absolute prediction error (MAPE), and the percentages of PE within 20% (F20) and 30% (F30) were calculated to assess the predictive performance of these models. In addition, previously built popPK model was included for comparison.

Results: Sixteen variables were selected as important covariates. Among ML models, the predictive performance of nonlinear-based ML models was superior to that of linear regression (MDPE: 3.27%, MAPE: 34.21%, F20: 30.63%, F30: 45.03%, R2 score: 0.68). The ML model built with the artificial neural network algorithm was considered the most suitable (MDPE: −0.039%, MAPE: 25.60%, F20: 39.35%, F30: 56.46%, R2 score: 0.75). Its performance was superior to that of the previously built popPK model (MDPE: 5.26%, MAPE: 29.22%, F20: 33.94%, F30: 51.22%, R2 score: 0.68). Furthermore, the application of the most suitable model and the popPK model in clinic showed that most dose regimen recommendations were reasonable.

Conclusion: The performance of these ML models indicate that a nonlinear relationship for covariates may help to improve model predictability. These results might facilitate the application of ML models in clinic, especially for patients with unstable status or during initial dose optimization.

A Systems-Based Analysis of Mono- and Combination Therapy for Carbapenem-Resistant Klebsiella pneumoniae Bloodstream Infections

Antimicrobial resistance is a global threat. As "proof-of-concept," we employed a system-based approach to identify patient, bacterial, and drug variables contributing to mortality in patients with carbapenem-resistant Klebsiella pneumoniae (CRKp) bloodstream infections exposed to colistin (COL) and ceftazidime-avibactam (CAZ/AVI) as mono- or combination therapies. Patients (n = 49) and CRKp isolates (n = 22) were part of the Consortium on Resistance Against Carbapenems in Klebsiella and other Enterobacteriaceae (CRACKLE-1), a multicenter, observational, prospective study of patients with carbapenem-resistant Enterobacterales (CRE) conducted between 2011 and 2016. Pharmacodynamic activity of mono- and combination drug concentrations was evaluated over 24 h using in vitro static time-kill assays. Bacterial growth and killing dynamics were estimated with a mechanism-based model. Random Forest was used to rank variables important for predicting 30-day mortality. Isolates exposed to COL+CAZ/AVI had enhanced early bacterial killing compared to CAZ/AVI alone and fewer incidences of regrowth compared to COL and CAZ/AVI. The mean coefficient of determination (R2) for the observed versus predicted bacterial counts was 0.86 (range: 0.75 - 0.95). Bacterial subpopulation susceptibilities and drug mechanistic synergy were essential to describe bacterial killing and growth dynamics. The combination of clinical (hypotension), bacterial (IncR plasmid, aadA2, and sul3) and drug (KC50) variables were most predictive of 30-day mortality. This proof-of-concept study combined clinical, bacterial, and drug variables in a unified model to evaluate clinical outcomes.

Machine learning advances the integration of covariates in population pharmacokinetic models: Valproic acid as an example

Background and Aim: Many studies associated with the combination of machine learning (ML) and pharmacometrics have appeared in recent years. ML can be used as an initial step for fast screening of covariates in population pharmacokinetic (popPK) models. The present study aimed to integrate covariates derived from different popPK models using ML.

Methods: Two published popPK models of valproic acid (VPA) in Chinese epileptic patients were used, where the population parameters were influenced by some covariates. Based on the covariates and a one-compartment model that describes the pharmacokinetics of VPA, a dataset was constructed using Monte Carlo simulation, to develop an XGBoost model to estimate the steady-state concentrations (Css) of VPA. We utilized SHapley Additive exPlanation (SHAP) values to interpret the prediction model, and calculated estimates of VPA exposure in four assumed scenarios involving different combinations of CYP2C19 genotypes and co-administered antiepileptic drugs. To develop an easy-to-use model in the clinic, we built a simplified model by using CYP2C19 genotypes and some noninvasive clinical parameters, and omitting several features that were infrequently measured or whose clinically available values were inaccurate, and verified it on our independent external dataset.

Results: After data preprocessing, the finally generated combined dataset was divided into a derivation cohort and a validation cohort (8:2). The XGBoost model was developed in the derivation cohort and yielded excellent performance in the validation cohort with a mean absolute error of 2.4 mg/L, root-mean-squared error of 3.3 mg/L, mean relative error of 0%, and percentages within ±20% of actual values of 98.85%. The SHAP analysis revealed that daily dose, time, CYP2C19*2 and/or *3 variants, albumin, body weight, single dose, and CYP2C19*1*1 genotype were the top seven confounding factors influencing the Css of VPA. Under the simulated dosage regimen of 500 mg/bid, the VPA exposure in patients who had CYP2C19*2 and/or *3 variants and no carbamazepine, phenytoin, or phenobarbital treatment, was approximately 1.74-fold compared to those with CYP2C19*1/*1 genotype and co-administered carbamazepine + phenytoin + phenobarbital. The feasibility of the simplified model was fully illustrated by its performance in our external dataset.

Conclusion: This study highlighted the bridging role of ML in big data and pharmacometrics, by integrating covariates derived from different popPK models.

An interpretable stacking ensemble learning framework based on multi-dimensional data for real-time prediction of drug concentration: The example of olanzapine

Background and Aim: Therapeutic drug monitoring (TDM) has evolved over the years as an important tool for personalized medicine. Nevertheless, some limitations are associated with traditional TDM. Emerging data-driven model forecasting [e.g., through machine learning (ML)-based approaches] has been used for individualized therapy. This study proposes an interpretable stacking-based ML framework to predict concentrations in real time after olanzapine (OLZ) treatment.

Methods: The TDM-OLZ dataset, consisting of 2,142 OLZ measurements and 472 features, was formed by collecting electronic health records during the TDM of 927 patients who had received OLZ treatment. We compared the performance of ML algorithms by using 10-fold cross-validation and the mean absolute error (MAE). The optimal subset of features was analyzed by a random forest-based sequential forward feature selection method in the context of the top five heterogeneous regressors as base models to develop a stacked ensemble regressor, which was then optimized via the grid search method. Its predictions were explained by using local interpretable model-agnostic explanations (LIME) and partial dependence plots (PDPs).

Results: A state-of-the-art stacking ensemble learning framework that integrates optimized extra trees, XGBoost, random forest, bagging, and gradient-boosting regressors was developed for nine selected features [i.e., daily dose (OLZ), gender_male, age, valproic acid_yes, ALT, K, BW, MONO#, and time of blood sampling after first administration]. It outperformed other base regressors that were considered, with an MAE of 0.064, R-square value of 0.5355, mean squared error of 0.0089, mean relative error of 13%, and ideal rate (the percentages of predicted TDM within ± 30% of actual TDM) of 63.40%. Predictions at the individual level were illustrated by LIME plots, whereas the global interpretation of associations between features and outcomes was illustrated by PDPs.

Conclusion: This study highlights the feasibility of the real-time estimation of drug concentrations by using stacking-based ML strategies without losing interpretability, thus facilitating model-informed precision dosing.

Machine learning approach identified clusters for patients with low cardiac output syndrome and outcomes after cardiac surgery

Background

Low cardiac output syndrome (LCOS) is the most serious physiological abnormality with high mortality for patients after cardiac surgery. This study aimed to explore the multidimensional data of clinical features and outcomes to provide individualized care for patients with LCOS.

Methods

The electronic medical information of the intensive care units (ICUs) was extracted from a tertiary hospital in South China. We included patients who were diagnosed with LCOS in the ICU database. We used the consensus clustering approach based on patient characteristics, laboratory data, and vital signs to identify LCOS subgroups. The consensus clustering method involves subsampling from a set of items, such as microarrays, and determines to cluster of specified cluster counts (k). The primary clinical outcome was in-hospital mortality and was compared between the clusters.

Results

A total of 1,205 patients were included and divided into three clusters. Cluster 1 (n = 443) was defined as the low-risk group [in-hospital mortality =10.1%, odds ratio (OR) = 1]. Cluster 2 (n = 396) was defined as the medium-risk group [in-hospital mortality =25.0%, OR = 2.96 (95% CI = 1.97–4.46)]. Cluster 3 (n = 366) was defined as the high-risk group [in-hospital mortality =39.2%, OR = 5.75 (95% CI = 3.9–8.5)].

Conclusion

Patients with LCOS after cardiac surgery could be divided into three clusters and had different outcomes.

Leveraging Predictive Pharmacometrics-Based Algorithms to Enhance Perinatal Care-Application to Neonatal Jaundice

The field of medicine is undergoing a fundamental change, transforming towards a modern data-driven patient-oriented approach. This paradigm shift also affects perinatal medicine as predictive algorithms and artificial intelligence are applied to enhance and individualize maternal, neonatal and perinatal care. Here, we introduce a pharmacometrics-based mathematical-statistical computer program (PMX-based algorithm) focusing on hyperbilirubinemia, a medical condition affecting half of all newborns. Independent datasets from two different centers consisting of total serum bilirubin measurements were utilized for model development (342 neonates, 1,478 bilirubin measurements) and validation (1,101 neonates, 3,081 bilirubin measurements), respectively. The mathematical-statistical structure of the PMX-based algorithm is a differential equation in the context of non-linear mixed effects modeling, together with Empirical Bayesian Estimation to predict bilirubin kinetics for a new patient. Several clinically relevant prediction scenarios were validated, i.e., prediction up to 24 h based on one bilirubin measurement, and prediction up to 48 h based on two bilirubin measurements. The PMX-based algorithm can be applied in two different clinical scenarios. First, bilirubin kinetics can be predicted up to 24 h based on one single bilirubin measurement with a median relative (absolute) prediction difference of 8.5% (median absolute prediction difference 17.4 μmol/l), and sensitivity and specificity of 95.7 and 96.3%, respectively. Second, bilirubin kinetics can be predicted up to 48 h based on two bilirubin measurements with a median relative (absolute) prediction difference of 9.2% (median absolute prediction difference 21.5 μmol/l), and sensitivity and specificity of 93.0 and 92.1%, respectively. In contrast to currently available nomogram-based static bilirubin stratification, the PMX-based algorithm presented here is a dynamic approach predicting individual bilirubin kinetics up to 48 h, an intelligent, predictive algorithm that can be incorporated in a clinical decision support tool. Such clinical decision support tools have the potential to benefit perinatal medicine facilitating personalized care of mothers and their born and unborn infants.

Contribution of machine learning to tumor growth inhibition modeling for hepatocellular carcinoma patients under Roblitinib (FGF401) drug treatment

Machine learning (ML) opens new perspectives in identifying predictive factors of efficacy among a large number of patients’ characteristics in oncology studies. The objective of this work was to combine ML with population pharmacokinetic/pharmacodynamic (PK/PD) modeling of tumor growth inhibition to understand the sources of variability between patients and therefore improve model predictions to support drug development decisions. Data from 127 patients with hepatocellular carcinoma enrolled in a phase I/II study evaluating once‐daily oral doses of the fibroblast growth factor receptor FGFR4 kinase inhibitor, Roblitinib (FGF401), were used. Roblitinib  PKs was best described by a two‐compartment model with a delayed zero‐order absorption and linear elimination. Clinical efficacy using the longitudinal sum of the longest lesion diameter data was described with a population PK/PD model of tumor growth inhibition including resistance to treatment. ML, applying elastic net modeling of time to progression data, was associated with cross‐validation, and allowed to derive a composite predictive risk score from a set of 75 patients’ baseline characteristics. The two approaches were combined by testing the inclusion of the continuous risk score as a covariate on PD model parameters. The score was found as a significant covariate on the resistance parameter and resulted in 19% reduction of its variability, and 32% variability reduction on the average dose for stasis. The final PK/PD model was used to simulate effect of patients’ characteristics on tumor growth inhibition profiles. The proposed methodology can be used to support drug development decisions, especially when large interpatient variability is observed.

Machine Learning and Pharmacometrics for Prediction of Pharmacokinetic Data: Differences, Similarities and Challenges Illustrated with Rifampicin

Pharmacometrics (PM) and machine learning (ML) are both valuable for drug development to characterize pharmacokinetics (PK) and pharmacodynamics (PD). Pharmacokinetic/pharmacodynamic (PKPD) analysis using PM provides mechanistic insight into biological processes but is time- and labor-intensive. In contrast, ML models are much quicker trained, but offer less mechanistic insights. The opportunity of using ML predictions of drug PK as input for a PKPD model could strongly accelerate analysis efforts. Here exemplified by rifampicin, a widely used antibiotic, we explore the ability of different ML algorithms to predict drug PK. Based on simulated data, we trained linear regressions (LASSO), Gradient Boosting Machines, XGBoost and Random Forest to predict the plasma concentration-time series and rifampicin area under the concentration-versus-time curve from 0–24 h (AUC_0–24h) after repeated dosing. XGBoost performed best for prediction of the entire PK series (R²: 0.84, root mean square error (RMSE): 6.9 mg/L, mean absolute error (MAE): 4.0 mg/L) for the scenario with the largest data size. For AUC_0–24h prediction, LASSO showed the highest performance (R²: 0.97, RMSE: 29.1 h·mg/L, MAE: 18.8 h·mg/L). Increasing the number of plasma concentrations per patient (0, 2 or 6 concentrations per occasion) improved model performance. For example, for AUC_0–24h prediction using LASSO, the R² was 0.41, 0.69 and 0.97 when using predictors only (no plasma concentrations), 2 or 6 plasma concentrations per occasion as input, respectively. Run times for the ML models ranged from 1.0 s to 8 min, while the run time for the PM model was more than 3 h. Furthermore, building a PM model is more time- and labor-intensive compared with ML. ML predictions of drug PK could thus be used as input into a PKPD model, enabling time-efficient analysis.

Artificial Neural Network vs. Pharmacometric Model for Population Prediction of Plasma Concentration in Real-World Data: A Case Study on Valproic Acid

We compared the predictive performance of an artificial neural network to traditional pharmacometric modeling for population prediction of plasma concentrations of valproate in real-world data. We included individuals aged 65 years or older with epilepsy who redeemed their first prescription of valproate after the diagnosis of epilepsy and had at least one valproate plasma concentration measured. A long short-term memory neural network (LSTM) was developed using the training data set to fit the LSTM and the test data set to validate the model. Predictions from the LSTM were compared with those obtained from the population predictions from a pharmacometric model by Birnbaum et al. which had the best predictive performance for population predictions of valproate concentrations in Danish databases. We used the cutoff of ± 20 mg/L of prediction error to define good predictions. A total of 1,252 individuals were included in the study. The LSTM fitted using the training data set had poor predictive performance in the test data set, but better than that of the pharmacometric model. The proportion of individuals with at least one predicted concentration within ± 20 mg/L of observed concentration was largest in case of the LSTM (64.4%, 95% confidence interval (CI): 58.4-70.2%) compared with the pharmacometric model by Birnbaum et al. (49.8%, 95% CI: 47.0-52.6%). LSTM shows better predictive performance to predict valproate plasma concentrations compared with a traditional pharmacometric model in the investigated setting with real-world data in older patients with epilepsy where information on exact timepoints for both dosing and plasma concentration measurement are missing.

Introduction of an artificial neural network–based method for concentration‐time predictions

Pharmacometrics and the application of population pharmacokinetic (PK) modeling play a crucial role in clinical pharmacology. These methods, which describe data with well‐defined equations and estimate physiologically interpretable parameters, have not changed substantially during the past decades. Although the methods have proven their usefulness, they are often resource intensive and require a high level of expertise. We investigated whether a method based on artificial neural networks (ANNs) may provide an alternative approach for the prediction of concentration‐time curve to supplement the gold standard methods. In this work, we used simulated data to overcome the requirement for a large clinical training data set, implemented a pharmacologically reasonable network architecture to improve extrapolation to different dosing schemes, and used transfer learning to quickly adapt the predictions to new patient groups. We demonstrate that ANNs are able to learn the shape of concentration‐time curves and make individual predictions based on a short sequence of PK measurements. Furthermore, an ANN trained on simulated data was applied to real clinical data and was demonstrated to extrapolate to different dosing schemes. We also adapted the ANN trained on simulated healthy subjects to simulated hepatic impaired patients through transfer learning. In summary, we demonstrate how ANNs could be leveraged in a PK workflow to efficiently make individual concentration‐time predictions, and we discuss the current limitations and advantages of such an ANN‐based method.

A Hybrid Model Associating Population Pharmacokinetics with Machine Learning: A Case Study with Iohexol Clearance Estimation

BACKGROUND

Maximum a posteriori Bayesian estimation (MAP-BE) based on a limited sampling strategy and a population pharmacokinetic model is frequently used to estimate pharmacokinetic parameters in individuals, however with some uncertainty (bias). Recent works have shown that the performance in individual estimation or pharmacokinetic parameters can be improved by combining population pharmacokinetic and machine learning algorithms.

OBJECTIVE

 The objective of this work was to investigate the use of a hybrid machine learning/population pharmacokinetic approach to improve individual iohexol clearance estimation.

METHODS

The reference iohexol clearance values were derived from 500 simulated profiles (samples collected between 0.1 and 24.7 h) using a population pharmacokinetic model we recently developed in Monolix and obtained using all the concentration timepoints available. Xgboost and glmnet algorithms able to predict the error of MAP-BE clearance estimates based on a limited sampling strategy (0.1 h, 1 h, and 9 h) versus reference values were developed in a training subset (75%) and were evaluated in a testing subset (25%) and in 36 real patients.

RESULTS

The MAP-BE limited sampling strategy estimated clearance was corrected by the machine learning predicted error leading to a decrease in root mean squared error by 29% and 24%, and in the percentage of profiles with the mean prediction error out of the ± 20% bias by 60% and 40% in the external validation dataset for the glmnet and Xgboost machine learning algorithms, respectively. These results were attributable to a decrease in the eta-shrinkage (shrinkage for a MAP-BE limited sampling strategy = 32.4%, glmnet = 18.2%, and Xgboost = 19.4% in the external dataset).

CONCLUSIONS

 In conclusion, this hybrid algorithm represents a significant improvement in comparison to MAP-BE estimation alone.

Artificial Intelligence in Infection Management in the ICU

This article is one of ten reviews selected from the Annual Update in Intensive Care and Emergency Medicine 2022. Other selected articles can be found online at https://www.biomedcentral.com/collections/annualupdate2022. Further information about the Annual Update in Intensive Care and Emergency Medicine is available from https://link.springer.com/bookseries/8901.

Pharmacometric modeling and machine learning analyses of prognostic and predictive factors in the JAVELIN Gastric 100 phase III trial of avelumab

Avelumab (anti–PD‐L1) is an approved anticancer treatment for several indications. The JAVELIN Gastric 100 phase III trial did not meet its primary objective of demonstrating superior overall survival (OS) with avelumab maintenance versus continued chemotherapy in patients with advanced gastric cancer/gastroesophageal junction cancer; however, the OS rate was numerically higher with avelumab at timepoints after 12 months. Machine learning (random forests, SIDEScreen, and variable‐importance assessments) was used to build models to identify prognostic/predictive factors associated with long‐term OS and tumor growth dynamics (TGDs). Baseline, re‐baseline, and longitudinal variables were evaluated as covariates in a parametric time‐to‐event model for OS and Gompertzian population model for TGD. The final OS model incorporated a treatment effect on the log‐logistic shape parameter but did not identify a treatment effect on OS or TGD. Variables identified as prognostic for longer OS included older age; higher gamma‐glutamyl transferase (GGT) or albumin; absence of peritoneal carcinomatosis; lower neutrophil‐lymphocyte ratio, lactate dehydrogenase, or C‐reactive protein (CRP); response to induction chemotherapy; and Eastern Cooperative Oncology Group performance status of 0. Among baseline and time‐varying covariates, the largest effects were found for GGT and CRP, respectively. Liver metastasis at re‐baseline predicted higher tumor growth. Tumor size after induction chemotherapy was associated with number of metastatic sites and stable disease (vs. response). Asian region did not impact OS or TGD. Overall, an innovative workflow supporting pharmacometric modeling of OS and TGD was established. Consistent with the primary trial analysis, no treatment effect was identified. However, potential prognostic factors were identified.

Pharmacometrics meets statistics-A synergy for modern drug development

Modern drug development problems are very complex and require integration of various scientific fields. Traditionally, statistical methods have been the primary tool for design and analysis of clinical trials. Increasingly, pharmacometric approaches using physiology-based drug and disease models are applied in this context. In this paper, we show that statistics and pharmacometrics have more in common than what keeps them apart, and collectively, the synergy from these two quantitative disciplines can provide greater advances in clinical research and development, resulting in novel and more effective medicines to patients with medical need.

Drug Clearance in Neonates: A Combination of Population Pharmacokinetic Modelling and Machine Learning Approaches to Improve Individual Prediction

BACKGROUND

Population pharmacokinetic evaluations have been widely used in neonatal pharmacokinetic studies, while machine learning has become a popular approach to solving complex problems in the current era of big data.

OBJECTIVE

The aim of this proof-of-concept study was to evaluate whether combining population pharmacokinetic and machine learning approaches could provide a more accurate prediction of the clearance of renally eliminated drugs in individual neonates.

METHODS

Six drugs that are primarily eliminated by the kidneys were selected (vancomycin, latamoxef, cefepime, azlocillin, ceftazidime, and amoxicillin) as 'proof of concept' compounds. Individual estimates of clearance obtained from population pharmacokinetic models were used as reference clearances, and diverse machine learning methods and nested cross-validation were adopted and evaluated against these reference clearances. The predictive performance of these combined methods was compared with the performance of two other predictive methods: a covariate-based maturation model and a postmenstrual age and body weight scaling model. Relative error was used to evaluate the different methods.

RESULTS

The extra tree regressor was selected as the best-fit machine learning method. Using the combined method, more than 95% of predictions for all six drugs had a relative error of < 50% and the mean relative error was reduced by an average of 44.3% and 71.3% compared with the other two predictive methods.

CONCLUSION

A combined population pharmacokinetic and machine learning approach provided improved predictions of individual clearances of renally cleared drugs in neonates. For a new patient treated in clinical practice, individual clearance can be predicted a priori using our model code combined with demographic data.

Fast screening of covariates in population models empowered by machine learning

One of the objectives of Pharmacometry (PMX) population modeling is the identification of significant and clinically relevant relationships between parameters and covariates. Here, we demonstrate how this complex selection task could benefit from supervised learning algorithms using importance scores. We compare various classical methods with three machine learning (ML) methods applied to NONMEM empirical Bayes estimates: random forest, neural networks (NNs), and support vector regression (SVR). The performance of the ML models is assessed using receiver operating characteristic (ROC) curves. The F1 score, which measures test accuracy, is used to compare ML and PMX approaches. Methods are applied to different scenarios of covariate influence based on simulated pharmacokinetics data. ML achieved similar or better F1 scores than stepwise covariate modeling (SCM) and conditional sampling for stepwise approach based on correlation tests (COSSAC). Correlations between covariates and the number of false covariates does not affect the performance of any method, but effect size has an impact. Methods are not equivalent with respect to computational speed; SCM is 30 and 100-times slower than NN and SVR, respectively. The results are validated in an additional scenario involving 100 covariates. Taken together, the results indicate that ML methods can greatly increase the efficiency of population covariate model building in the case of large datasets or complex models that require long run-times. This can provide fast initial covariate screening, which can be followed by more conventional PMX approaches to assess the clinical relevance of selected covariates and build the final model.

Machine learning to predict high-dose methotrexate-related neutropenia and fever in children with B-cell acute lymphoblastic leukemia

Methotrexate (MTX), an antimetabolite for the treatment of leukemia, could cause neutropenia and subsequently fever, which might lead to treatment delay and affect prognosis. Here, we aimed to predict neutropenia and fever related to high-dose MTX using artificial intelligence. This study included 139 pediatric patients newly diagnosed with standard- or intermediate risk B-cell acute lymphoblastic leukemia. Fifty-seven SNPs of 16 genes were genotyped. Univariate and multivariate analysis were used to select SNPs and clinical covariates for model developing. Five machine learning algorithms combined with four resampling techniques were used to build optimal predictive model. The combination of random forest with adaptive synthetic appeared to be the best model for neutropenia (sensitivity = 0.935, specificity = 0.920, AUC = 0.927) and performed best for fever (sensitivity = 0.818, specificity = 0.924, AUC = 0.870). By machine learning, we have developed and validated comprehensive models to predict the risk of neutropenia and fever. Such models may be helpful for medical oncologists in quick decision-making.

Pharm-AutoML: An open-source, end-to-end automated machine learning package for clinical outcome prediction

Although there is increased interest in utilizing machine learning (ML) to support drug development, technical hurdles associated with complex algorithms have limited widespread adoption. In response, we have developed Pharm‐AutoML, an open‐source Python package that enables users to automate the construction of ML models and predict clinical outcomes, especially in the context of pharmacological interventions. In particular, our approach streamlines tedious steps within the ML workflow, including data preprocessing, model tuning, model selection, results analysis, and model interpretation. Moreover, our open‐source package helps to identify the most predictive ML pipeline among defined search spaces by selecting the best data preprocessing strategy and tuning the ML model hyperparameters. The package currently supports multiclass classification tasks, and additional functions are being developed. Using a set of five publicly available biomedical datasets, we show that our Pharm‐AutoML outperforms other ML frameworks, including H2O with default settings, by demonstrating improved predictive accuracy of classification. We further illustrate how model interpretation methods can be utilized to help explain the fine‐tuned ML pipeline to end users. Pharm‐AutoML provides both novice and expert users improved clinical predictions and scientific insights.

Machine learning in pharmacometrics: Opportunities and challenges

The explosive growth in medical devices, imaging and diagnostics, computing, and communication and information technologies in drug development and healthcare has created an ever-expanding data landscape that the pharmacometrics (PMX) research community must now traverse. The tools of machine learning (ML) have emerged as a powerful computational approach in other data-rich disciplines but its effective utilization in the pharmaceutical sciences and PMX modelling is in its infancy. ML-based methods can complement PMX modelling by enabling the information in diverse sources of big data, e.g. population-based public databases and disease-specific clinical registries, to be harnessed because they are capable of efficiently identifying salient variables associated with outcomes and delineating their interdependencies. ML algorithms are computationally efficient, have strong predictive capabilities and can enable learning in the big data setting. ML algorithms can be viewed as providing a computational bridge from big data to complement PMX modelling. This review provides an overview of the strengths and weaknesses of ML approaches vis-à-vis population methods, assesses current research into ML applications in the pharmaceutical sciences and provides perspective for potential opportunities and strategies for the successful integration and utilization of ML in PMX.

Early predictions of response and survival from a tumor dynamics model in patients with recurrent, metastatic head and neck squamous cell carcinoma treated with immunotherapy

We developed and evaluated a method for making early predictions of best overall response (BOR) and overall survival at 6 months (OS6) in patients with cancer treated with immunotherapy. This method combines machine learning with modeling of longitudinal tumor size data. We applied our method to data from durvalumab‐exposed patients with recurrent/metastatic head and neck cancer. A fivefold cross‐validation was used for model selection. Independent trial data, with various degrees of data truncation, were used for model validation. Mean classification error rates (90% confidence intervals [CIs]) from cross‐validation were 5.99% (90% CI 2.98%–7.50%) for BOR and 19.8% (90% CI 15.8%–39.3%) for OS6. During model validation, the area under the receiver operating characteristic curves was preserved for BOR (0.97, 0.97, and 0.94) and OS6 (0.85, 0.84, and 0.82) at 24, 18, and 12 weeks, respectively. These results suggest our method predicts trial outcomes accurately from early data and could be used to aid drug development.

Using Machine Learning to Predict the Diagnosis, Management and Severity of Pediatric Appendicitis

Background: Given the absence of consolidated and standardized international guidelines for managing pediatric appendicitis and the few strictly data-driven studies in this specific, we investigated the use of machine learning (ML) classifiers for predicting the diagnosis, management and severity of appendicitis in children. Materials and Methods: Predictive models were developed and validated on a dataset acquired from 430 children and adolescents aged 0-18 years, based on a range of information encompassing history, clinical examination, laboratory parameters, and abdominal ultrasonography. Logistic regression, random forests, and gradient boosting machines were used for predicting the three target variables. Results: A random forest classifier achieved areas under the precision-recall curve of 0.94, 0.92, and 0.70, respectively, for the diagnosis, management, and severity of appendicitis. We identified smaller subsets of 6, 17, and 18 predictors for each of targets that sufficed to achieve the same performance as the model based on the full set of 38 variables. We used these findings to develop the user-friendly online Appendicitis Prediction Tool for children with suspected appendicitis. Discussion: This pilot study considered the most extensive set of predictor and target variables to date and is the first to simultaneously predict all three targets in children: diagnosis, management, and severity. Moreover, this study presents the first ML model for appendicitis that was deployed as an open access easy-to-use online tool. Conclusion: ML algorithms help to overcome the diagnostic and management challenges posed by appendicitis in children and pave the way toward a more personalized approach to medical decision-making. Further validation studies are needed to develop a finished clinical decision support system.

Application of Machine Learning for Tumor Growth Inhibition - Overall Survival Modeling Platform

Machine learning (ML) was used to leverage tumor growth inhibition (TGI) metrics to characterize the relationship with overall survival (OS) as a novel approach and to compare with traditional TGI‐OS modeling methods. Historical dataset from a phase III non‐small cell lung cancer study (OAK, atezolizumab vs. docetaxel, N = 668) was used. ML methods support the validity of TGI metrics in predicting OS. With lasso, the best model with TGI metrics outperforms the best model without TGI metrics. Boosting was the best linear ML method for this dataset with reduced estimation bias and lowest Brier score, suggesting better prediction accuracy. Random forest did not outperform linear ML methods despite hyperparameter optimization. Kernel machine was marginally the best nonlinear ML method for this dataset and uncovered nonlinear and interaction effects. Nonlinear ML may improve prediction by capturing nonlinear effects and covariate interactions, but its predictive performance and value need further evaluation with larger datasets.

From Real-World Patient Data to Individualized Treatment Effects Using Machine Learning: Current and Future Methods to Address Underlying Challenges

Clinical decision making needs to be supported by evidence that treatments are beneficial to individual patients. Although randomized control trials (RCTs) are the gold standard for testing and introducing new drugs, due to the focus on specific questions with respect to establishing efficacy and safety vs. standard treatment, they do not provide a full characterization of the heterogeneity in the final intended treatment population. Conversely, real-world observational data, such as electronic health records (EHRs), contain large amounts of clinical information about heterogeneous patients and their response to treatments. In this paper, we introduce the main opportunities and challenges in using observational data for training machine learning methods to estimate individualized treatment effects and make treatment recommendations. We describe the modeling choices of the state-of-the-art machine learning methods for causal inference, developed for estimating treatment effects both in the cross-section and longitudinal settings. Additionally, we highlight future research directions that could lead to achieving the full potential of leveraging EHRs and machine learning for making individualized treatment recommendations. We also discuss how experimental data from RCTs and Pharmacometric and Quantitative Systems Pharmacology approaches can be used to not only improve machine learning methods, but also provide ways for validating them. These future research directions will require us to collaborate across the scientific disciplines to incorporate models based on RCTs and known disease processes, physiology, and pharmacology into these machine learning models based on EHRs to fully optimize the opportunity these data present.