Sepsis Detection Exploiting Biomarker Analysis with Deep Neural Networks

Sepsis Detection Exploiting Biomarker Analysis with Deep Neural Networks Dimitrios Spanos1, Nikolaos Passalis2, Dimosthenis Spasopoulos3, Evangelia Chatzianagnostou3, Juan Carlos Ruiz Rodr´ıguez4, Juan Jose Gonzalez Lopez4, Laura M. Lechuga5, M.-Carmen Est´evez5, Nikos Pleros3, and Anastasios Tefas1 1Computational Intelligence and Deep Learning Group, AIIA Lab, Dept. of Informatics 2Dept. of Chemical Engineering 3Wireless and Photonic Systems and Networks Group, Dept. of Informatics Aristotle University of Thessaloniki, Thessaloniki, Greece 4Hospital Universitari Vall d’Hebron, Barcelona, Spain 5Nanobiosensors and Bioanalytical Applications Group (NanoB2A), Catalan Institute of Nanoscience and Nanotechnology (ICN2), CSIC, CIBER-BBN and BIST, Campus UAB, Bellaterra, 08193, Barcelona, Spain E-mails: [email protected], {passalis, dspa}@auth.gr, evac[email protected], {juancarlos.ruiz, juanjo.gonzalez}@vallhebron.cat, {laura.lechuga, mcarmen.estevez}@icn2.cat, {pleros, tefas}@csd.auth.gr Abstract Machine learning has found widespread application in many domains, including medical sciences, offering promising solutions to complex healthcare challenges using cyber-physical systems (CPSs). Although mortality rates for sepsis and septic shock have declined in recent years, these conditions remain critical and are often overlooked, necessitating advanced predictive modeling to enable prompt and accurate diagnosis. Providing informative indicators for physicians and enabling early sepsis diagnosis are critical for improving patient outcomes, even when minimal information is accessible. In this paper, we present a deep learning (DL) pipeline for early sepsis detection when only a limited number of biomarkers is available. Furthermore, Preprint submitted to ICPS October 16, 2025 we also introduce a domain-guided data refinement pipeline that combines prior domain knowledge with the predictions of a DL model to guide the data gathering and refinement process. We demonstrate the effectiveness of the proposed approach through extensive experiments across a wide range of settings and tasks. Our findings show the potential of deep neural networks as a helpful tool for sepsis diagnosis, enabling timely interventions and better patient care. 1. Introduction Machine learning has seen successful application in various fields, including medical sciences, where it has powered advancements in diagnostics, hospital cyber-physical systems and patient monitoring [1] [2]. In the case of sepsis and septic shock [3], while mortality rates have decreased in recent years [4], these conditions remain critical challenges due to their rapid onset, potential long-term effects [5], and the complexity of early detection. Despite their importance, research into using deep learning for sepsis diagnosis has often been overlooked, especially in scenarios where quick predictions are required and only a limited set of biomarkers is available. The motivation for this work comes from the challenges faced by domain experts, such as physicians, in creating datasets that aid in the training process of deep neural networks and lead to optimal performance. Clinical data, in its raw form, is rarely ready for use in machine learning applications [6], often requiring extensive preprocessing and refinement. This process can be time-consuming and error-prone, which can be especially critical in applications like sepsis diagnosis, where inconsistencies can have serious consequences. Addressing this challenge calls for methodologies that effectively integrate domain expertise with deep learning learning capabilities to streamline the refinement process. These challenges bring us to the core question of this paper: Can a deep neural network help domain experts refine clinical data in a meaningful way that enhances model performance? To tackle this question, we propose a methodology called Domain-Guided Data Refinement (DGDR). DGDR combines prior knowledge from domain experts with deep learning(DL) models/deep neural networks (DNN) to provide increased explainability and guide the data refinement process. Specifically, the proposed method enables a DNN to assign feature importance scores to biomarkers, offering 2 clear insights into which biomarkers are most critical in the decision-making process. Using these insights, domain experts can validate and refine the dataset by reexamining key samples against established guidelines, ensuring high-quality data construction. The main contribution of this paper is a novel methodology that integrates domain expertise and deep neural networks to streamline the data refinement process, improving performance in sepsis diagnosis and providing explainability through feature importance scores, allowing experts to understand which biomarkers drive the model’s predictions. Through extensive experiments, we demonstrate the superiority of our approach over baseline machine learning methods in various scenarios, showcasing its adaptability and reliability. The rest of the paper is organized as follows. The methodology is described in Section II. The experimental setup and evaluation of the proposed method are presented in Section IV. Finally, discussions are made in Section IV and conclusions are drawn in Section V. 2. Methodology 2.1. Background and Notation Let xi∈RDdenote an input sample, where Dis the dimensionality of the input space, and yi∈ {0,1}Cdenote its one-hot label vector such that PC c=1 yi,c = 1 in a set of i= 1, . . . , N training samples spanning over C distinct classes. The deep neural network is composed of Lfully connected layers, with the l-th layer containing Nlneurons, where l∈ {1,2, . . . , L}and the final output layer contains Cneurons corresponding to the number of classes. The network f(·)∈RCis trained to map input samples xito their corresponding one-hot labels yi, so that the output of the network provides predictions for in-domain samples. The activation function is denoted as σ, where σ(·) is typically a ReLU activation function for hidden layers and a softmax activation for the output layer. The output of the l-th layer can be calculated as: zl=Wlσ(zl−1)+bl,for l= 1, . . . , L, (1) where Wl∈RNl×Nl−1is the weight matrix at the l-th layer, and bl∈RNlis the bias vector at the l-th layer. For the first layer, z0=xi. 3 The final output layer ˆ yi∈RCrepresents the predicted class probabilities, calculated using the softmax function as follows: ˆyi,c =exp(z(c) L) PC c′=1 exp(z(c′) L),(2) which ensures that the sum of the predicted probabilities across all classes equals 1, i.e., PC c=1 ˆyi,c = 1. The network is trained by minimizing the cross-entropy loss between the predicted class probabilities ˆ yiand the true one-hot label vector yi: Lclass =− C X c=1 yi,c log(ˆyi,c),(3) where yi,c is the target label for class cand ˆyi,c is the predicted probability for class c. The network is also regularized with L2regularization on the weights in order to avoid overfitting: Lreg =λ L X l=1 ∥Wl∥2 2+ L X l=1 ∥bl∥2 2!(4) where λis a hyperparameter controlling the strength of the regularization, and ∥Wl∥2 2is the squared L2norm of the weight matrix Wlat layer l. Thus, the final learning objective Lused to train the network is the sum of the classification loss and the regularization loss: L=Lclass +Lreg.(5) The weights W1,W2,...,WLand the biases b1,b2,...,bLare updated using gradient descent to minimize the total loss function, ensuring that the network learns to accurately map input samples to their corresponding labels. 2.2. Domain-Guided Data Refinement We propose a domain-guided data refinement approach to improve the quality of a dataset Di(initially Di=D1). The steps involved are as follows: 1. First, we train a deep neural network on the dataset Dito predict labels for all samples. Let the input sample xi∈ Dihave a corresponding predicted label ˆyigenerated by the model. The prediction for each sample is computed as: ˆyi=f(xi),∀xi∈ Di.(6) 4 2. Next, we calculate feature importance scores for each feature of each sample in Di: si,j = ∂L(ˆyi) ∂xi,j  ,(7) where si,j denotes the importance score for the j-th feature of the ith sample, L(ˆyi) is the loss function, and xi,j is the value of the j-th feature of the i-th sample. The normalized feature importance scores snorm i,j are computed as: snorm i,j =si,j PD j=1 si,j .(8) 3. For each incorrectly classified sample xi, we focus on the top Fmost important features, based on the normalized feature importance scores. For these top Ffeatures, we check whether their values comply with predefined domain-specific guidelines. If any of the top Ffeatures of a sample do not meet the guidelines, the sample is flagged as inconsistent with domain knowledge (we discuss later how this check can be performed in a practical and efficient way). This can be expressed as: Flag(xi,j, li) = (1,if fD b(xi,j, l)>0 0,otherwise,(9) where fD b(xi,j, l) encoded the prior domain knowledge and represents a function that checks whether a sample classified as liis biologically/physically plausible when biomarker bhas the value of feature xi,j. 4. After consulting with domain experts, the flagged samples are reviewed, and corrections are made based on domain knowledge. This can result to conclusions on which samples need correcting and whether or not new features or samples should be added to a new dataset Di+1. The data refinement process can be repeated iteratively by retraining the model on the updated dataset Di+1, recalculating feature importance scores, and refining the data further. Note that domain knowledge is encoded in a set of functions fD b(x, l), where bdenotes a biomarker, xthe corresponding measurements and lthe 5 Figure 1: Sepsis guidelines for the employed biomarkers C-reactive protein (CRP), Interleukin-6 (IL-6), Procalcitonin (PCT) and Mid-regional pro-adrenomedullin (MRproADM) used to define prior domain knowledge and mark potentially anomalous samples for review by domain experts. label predicted by a predictive model. Each fD b(x, l) is defined as follows: fD b(x, l) =      1,if the measurement xof biomarker b cannot correspond to label l, 0,otherwise, (10) Note that the aim of fD b(x, l) is not to reclassify the samples but rather to identify cases where the prediction provided by a predictive model is inconsistent with prior knowledge. In this work, we employed the sepsis guidelines for the used biomarkers, as provided by domain experts, and summarized in Fig. 1. 6 3. Experiments In this section, we present the experimental evaluation of the proposed approach. First, we describe the datasets used in this study, including the initial dataset and the updated version obtained through the proposed DGDR process. Next, we outline the experimental setup and the evaluation metrics employed. Finally, we provide an extensive evaluation across various settings and tasks to demonstrate the effectiveness of the developed model for sepsis detection. 3.1. Datasets First Version (D1)is a dataset created to enable training machine learning for sepsis detection. Each collected sample was classified into one of the following categories: “Non-Infectious Systemic Inflammatory Response Syndrome (SIRS)” (175 samples), “Sepsis” (242 samples) and “Septic Shock” (331 samples). Several data from 785 patients were collected, providing basic information such as Sequential Organ Failure Assessment (SOFA) scores, blood culture results and microorganism names. However, results like SOFA scores and blood culture findings are typically available only several hours after a patient’s admission and do not enable early sepsis detection. The biomarkers available in D1and used for the corresponding experiments are: •C-reactive protein (CRP) [7]: is a marker of systemic inflammation that increases in response to infection or tissue injury, aiding in the early detection of sepsis. •Interleukin-6 (IL-6) [8]: is a pro-inflammatory cytokine that plays a critical role in the immune response and is often elevated in patients with sepsis. •Procalcitonin (PCT) [9]: is a peptide precursor of calcitonin, widely used as a biomarker for bacterial infections and sepsis because of its rapid and specific rise in response to systemic infections. Second Version (D2)is the updated version of the dataset D1, created after applying the proposed DGDR process and consulting with domain experts and physicians, for the purpose of assisting predictive models in the sepsis detection task. The labels were assigned using three different ways, proposed by domain experts after reviewed the predictions of DL models 7 on D1. The first consists of labels assigned internally by the hospital: “Infection without SIRS” (52 samples), “No infection” (31 samples), “SIRS” (130 samples), “Septic shock” (404 samples), and “Severe Sepsis” (331 samples). The second one uses the Sepsis-2 definition [10] with labels “Infection without SIRS” (52 samples), “No infection-no SIRS” (31 samples), “No infection-SIRS” (130 samples), “Septic shock” (404 samples), and “Severe Sepsis” (331 samples), and the last one uses the Sepsis-3 definition [3] with labels “Complicated infection” (45 samples), “Infection” (11 samples), “No infection” (159 samples), “Sepsis” (489 samples), “Septic Shock” (232 samples). In addition, more patient samples were collected, increasing the total to 949. Furthermore, domain experts identified instances where the included biomarkers could not be easily correlated with the actual category of a sample. To address these cases more effectively, two additional biomarkers were incorporated into the updated dataset: •Lactate [11]: is a byproduct of anaerobic metabolism and a key indicator of tissue hypoxia. Elevated lactate levels are associated with organ perfusion and increased mortality in sepsis. •Mid-regional pro-adrenomedullin (MR-proADM) [12]: is a biomarker reflecting endothelial dysfunction and vascular integrity, providing prognostic information on severity and progression of sepsis. This biomarker is available for only 585 out of the 949 patients. 3.2. Experimental Setup Hyper parameters For all conducted experiments we use a neural network architecture with three layers comprising 64, 32 and Cneurons respectively, where Cindicates the number of classes of the classification task. Furthermore, we use the ReLU activation function and the Adam [13] optimizer. Hyper-parameters such as learning rate, batch size and epochs were tuned with a simple grid search. For all conducted experiments, 30% of the available data was withheld as testing data in each run. Baselines For the purpose of comparing the proposed method with baseline traditional machine learning models, we use the scikit-learn [14] implementation of: Decision Trees, KNN (k=5), Logistic Regression, Support Vector Machine (SVM) and Ridge Regression. Evaluation Measures Accuracy measures the proportion of correctly classified samples, defined as Accuracy = TP+TN TP+FP+FN+TN, where TP (True Positives) represents correctly predicted positive samples, TN (True Negatives) 8 Table 1: Sepsis prediction evaluation on D1using CRP, PCT and IL6 biomarkers. Each sample is classified to one of the following classes: “Non-Infectious SIRS” , “Sepsis” and “Septic Shock”. Mean and standard deviation of accuracy and F-score are shown, calculated over 10 runs. Model Accuracy F-score Decision Tree 0.4231 ±0.0164 0.4233 ±0.0160 KNN 0.4333 ±0.0344 0.4365 ±0.0312 Logistic Regression 0.4898 ±0.0327 0.4724 ±0.0484 SVM 0.4769 ±0.0391 0.4604 ±0.0537 Ridge Classifier 0.4747 ±0.0365 0.4324 ±0.0630 DNN 0.5040 ±0.0332 0.4900 ±0.0312 Table 2: Sepsis prediction evaluation on D1using CRP, PCT and IL6 biomarkers. Each sample is classified to one of the following classes: “Non-Infectious SIRS” and “Sepsis or Septic Shock”. Mean and standard deviation of accuracy and F-score are shown, calculated over 10 runs. Model Accuracy F-score Decision Tree 0.6964 ±0.0264 0.6989 ±0.0320 KNN 0.7364 ±0.0228 0.7189 ±0.0318 Logistic Regression 0.7604 ±0.0261 0.6572 ±0.0353 SVM 0.7604 ±0.0261 0.6572 ±0.0353 Ridge Classifier 0.7604 ±0.0261 0.6572 ±0.0353 DNN 0.7711 ±0.0268 0.7049 ±0.0550 denotes correctly predicted negative samples, FP (False Positives) are negative samples incorrectly predicted as positive, and FN (False Negatives) are positive samples incorrectly predicted as negative. The F-score, given by F1 = 2 ·Precision·Recall Precision+Recall , balances precision, Precision = TP TP+FP, and recall, Recall = TP TP+FN. The F-score is particularly important in predictive models for healthcare, as a high number of FN could result in the failure to identify serious medical conditions, potentially leading to severe consequences. 3.3. D1Dataset Evaluation Severity of Sepsis Prediction In Table 1 we show results of baselines and the DNN on D1where C= 3 and labels “Non-Infectious SIRS”, “Sepsis” and 9 [2] L. Wang, S. Xi, W. Liu, Q. Zhou, Duty cycle optimization for blood pressure sensors in wireless body area networks based on reinforcement learning, in: 2021 4th IEEE International Conference on Industrial CyberPhysical Systems (ICPS), IEEE, 2021, pp. 799–804. [3] M. Singer, C. S. Deutschman, C. W. Seymour, M. Shankar-Hari, D. Annane, M. Bauer, R. Bellomo, G. R. Bernard, J.-D. Chiche, C. M. Coopersmith, et al., The third international consensus definitions for sepsis and septic shock (sepsis-3), Jama 315 (8) (2016) 801–810. [4] E. K. Stevenson, A. R. Rubenstein, G. T. Radin, R. S. Wiener, A. J. Walkey, Two decades of mortality trends among patients with severe sepsis: a comparative meta-analysis, Critical care medicine 42 (3) (2014) 625–631. [5] B. D. Winters, M. Eberlein, J. Leung, D. M. Needham, P. J. Pronovost, J. E. Sevransky, Long-term mortality and quality of life in sepsis: a systematic review, Critical care medicine 38 (5) (2010) 1276–1283. [6] R. Miotto, F. Wang, S. Wang, X. Jiang, J. T. Dudley, Deep learning for healthcare: review, opportunities and challenges, Briefings in bioinformatics 19 (6) (2018) 1236–1246. [7] G. P. Castelli, C. Pognani, M. Meisner, A. Stuani, D. Bellomi, L. Sgarbi, Procalcitonin and c-reactive protein during systemic inflammatory response syndrome, sepsis and organ dysfunction, Critical care 8 (2004) 1–9. [8] D. M. Franco, I. Arevalo-Rodriguez, M. R. i Figuls, J. Zamora, Interleukin-6 for diagnosis of sepsis in critically ill adult patients, The Cochrane Database of Systematic Reviews 2015 (7) (2015). [9] A. L. Vijayan, n. Vanimaya, S. Ravindran, R. Saikant, S. Lakshmi, R. Kartik, Procalcitonin: a promising diagnostic marker for sepsis and antibiotic therapy, Journal of intensive care 5 (2017) 1–7. [10] I. Matot, C. L. Sprung, Definition of sepsis, Intensive care medicine 27 (14) (2001) S3–S9. [11] R. Gauer, D. Forbes, N. Boyer, Sepsis: diagnosis and management, American family physician 101 (7) (2020) 409–418. 16 [12] U. ¨ Onal, F. Valenzuela-S´anchez, K. E. Vandana, J. Rello, Mid-regional pro-adrenomedullin (mr-proadm) as a biomarker for sepsis and septic shock: narrative review, in: Healthcare, Vol. 6, MDPI, 2018, p. 110. [13] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014). [14] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Scikit-learn: Machine learning in python, Journal of Machine Learning Research 12 (2011) 2825–2830. 17