Unraveling the impact of therapeutic drug monitoring via machine learning for patients with sepsis

Abstract

Clinical studies investigating the benefits of beta-lactam therapeutic drug monitoring (TDM) among critically ill patients are hindered by small patient groups, variability between studies, patient heterogeneity, and inadequate use of TDM. Accordingly, definitive conclusions regarding the efficacy of TDM remain elusive. To address these challenges, we propose an innovative approach that leverages data-driven methods to unveil the concealed connections between therapy effectiveness and patient data, collected through a randomized controlled trial (DRKS00011159; 10th October 2016). Our findings reveal that machine learning algorithms can successfully identify informative features that distinguish between healthy and sick states. These hold promise as potential markers for disease classification and severity stratification, as well as offering a continuous and data-driven "multidimensional" Sequential Organ Failure Assessment (SOFA) score. The positive impact of TDM on patient recovery rates is demonstrated by unraveling the intricate connections between therapy effectiveness and clinically relevant data via machine learning.

Keywords: Mahalanobis distance; SOFA score; beta-lactam antibiotics; intensive care unit; machine learning; mathematical similarity; piperacillin/tazobactam; sepsis; septic shock; state space approach; therapeutic drug monitoring.

Graphical abstract

Figure 1

Quantifying effect of TDM by a similarity-based state-space approach (A) Data processing and feature selection pipeline for patient status analysis using the in-house data preparation code. The heterogeneous medical database collected from hospitals is transformed into digital patient cards, followed by the selection of the top 28 representative features through feature engineering. A similarity-based state tracking approach is employed to compare the impact of TDM on patient status. The integration of biosensors for frequent sampling and enhanced drug dosage control is proposed to complete the loop and further optimize patient care. (B) Feature selection workflow utilizing genetic algorithm (GA) implementation. The process involves leaving out 10 patients for generalizability testing, train/test split for feature selection evaluation, feature scaling/transformation, and iterative refinement with GA with cross-validation. The final feature set is determined through a frequency analysis of 100 repetitions. (C) Visualization of traveled distance analogy in a 2D feature space to assess patient state dissimilarity. The blue and orange points represent TDM and control group patients, respectively. The distance to the reference health state (d_i) indicates the degree of dissimilarity from the “healthy” state. Patients in both groups start their recovery trajectory in a specific sub-space of the 2D state space and are expected to move toward the reference state over time. The rate of recovery is determined by how quickly the groups progress from their initial states to the reference “healthy town” of Sequential Organ Failure Assessment (SOFA) score of 1. The cumulative sum of Mahalanobis distances is calculated to quantify the difference between TDM and control groups’ proximity to the healthy zone for each day.

Figure 2

Data-driven assessment of TDM and the impact of TDM on patient state trajectory The underlying hypothesis of this study is that the medical data collected daily during the clinical study hold valuable information regarding the patients’ health states. By employing mathematical techniques of similarity, the gradual changes in patient states based on the distribution of their health states were quantified. (A and B) Mathematical similarity of patient groups on day 1 is demonstrated using two dimensionality reduction techniques: (A) t-distributed stochastic neighbor embedding (t-SNE) and (B) principal component analysis (PCA). Both t-SNE and PCA embeddings indicate that patient states were homogenously distributed on the day of admission, validating the random TDM and control split. Quantitative comparison in high-dimensional feature space is given in the supplementary information. (C) The concept of “traveling to a healthier state,” which is evaluated by calculating the normalized Mahalanobis distance between the patient’s health status on each day and the reference state. (D) Randomly sampled individual patient trajectories from both the control and TDM groups, revealing the disparity in mathematical distance toward the SOFA 1 state.

Figure 3

Effect of TDM on patient recovery trajectories (A) The number of people left the study alive in both control and TDM groups. (B) The number of people left the study dead in both control and TDM groups. (C) Last recorded SOFA scores (median) for patients left the study alive in both the TDM and control groups. (D) The presence of distinct pathogen types (red) with the distribution of piperacillin-resistant pathogen (blue) in both the TDM and control groups on day 1 (following randomization). (E) The presence of distinct pathogen types (red) with the distribution of sepsis-causing pathogen (blue) in both the TDM and control groups on day 1 (following randomization). Pathogen distributions and the distribution of piperacillin-resistant and sepsis-causing pathogens exhibit similarities between the two groups, indicating that the patients in TDM and control sub-populations started the treatment in similar conditions (see Table S1).

Figure 4

Features considered relevant by the evolutionary feature selection algorithm Each bar denotes how many times a feature was picked by GA for SOFA score prediction. Features used in the patient state analysis are highlighted as dark blue. (A) Continuous features encompass various information (such as age, height, and body weight), laboratory (leukocyte count, hematocrit levels, and creatinine levels), drug-related information (concentration and infusion rate), and physiological measurements (such as breathing rate, body temperature, and mean arterial pressure). (B) Discrete features consist of yes/no questions and ordinal variables, such as the presence of metabolic acidosis, renal dysfunction, or the need for renal replacement. (C) Whether a pathogen could be detected in the patient. Microbiology reports cover 36 different pathogens, including gram-positive and gram-negative bacteria, fungi, and other pathogens such as Chlamydia species; (D) whether a detected pathogen is resistant to piperacillin and (E) whether the pathogen type is responsible for the sepsis episode. See Table S1 for more details.