Multi-objective optimization of fracturing ball strength and corrosion rate with genetic algorithms and interpretable machine learning

Improving Clinical Decisions in IR: Interpretable Machine Learning Models for Predicting Ascites Improvement after Transjugular Intrahepatic Portosystemic Shunt Procedures

To prevent rapid degradation causing plugging failure in degradable magnesium alloy packer slips under high-temperature, high-pressure, corrosive conditions, a composite dual-coated magnesium alloy was developed. The coating (hereafter referred to as Black Composite Coating, BCC) utilized micro-arc oxidation (MAO) and spray coating. SEM, XRD, FTIR, polarization curves, and EIS characterized microstructure and corrosion performance pre- and post-modification. Results showed the MAO-treated surface formed a porous structure, increasing specific surface area and roughness. This enhanced base/auxiliary material adhesion and corrosion resistance. The added black composite coating significantly reduced corrosion current density and increased charge transfer resistance (Rct). MAO pores provided mechanical interlocking sites. The black coating's low porosity and sealing blocked corrosive media penetration. Compared to soluble magnesium alloys, the coating reduced corrosion current density by ~five orders of magnitude (from ~10−4 to ~10−9 A/cm2) and increased Rct by ~six orders (from ~102 to ~106 Ω·cm2). Corrosion rates were suppressed >91.8%, stabilizing between 0.1~1.0 mm/year. Per ISO 9223:2012, this moderate corrosion resistance (Class C3-C4) indicates suitability for mildly aggressive environments like indoor industrial areas, low-chloride rural zones, or oil wells. This technique enhances degradable magnesium alloy corrosion resistance downhole, providing valuable industrial guidance.

BACKGROUND Acute kidney injury (AKI) is a common adverse outcome following nephrectomy. The progression from AKI to acute kidney disease (AKD) and subsequently to chronic kidney disease (CKD) remains a concern; yet, the predictive mechanisms for these transitions are not fully understood. Interpretable machine learning (ML) models offer insights into how clinical features influence long-term renal function outcomes after nephrectomy, providing a more precise framework for identifying patients at risk and supporting improved clinical decision-making processes. OBJECTIVE This study aimed to (1) evaluate postnephrectomy rates of AKI, AKD, and CKD, analyzing long-term renal outcomes along different trajectories; (2) interpret AKD and CKD models using Shapley Additive Explanations values and Local Interpretable Model-Agnostic Explanations algorithm; and (3) develop a web-based tool for estimating AKD or CKD risk after nephrectomy. METHODS We conducted a retrospective cohort study involving patients who underwent nephrectomy between July 2012 and June 2019. Patient data were randomly split into training, validation, and test sets, maintaining a ratio of 76.5:8.5:15. Eight ML algorithms were used to construct predictive models for postoperative AKD and CKD. The performance of the best-performing models was assessed using various metrics. We used various Shapley Additive Explanations plots and Local Interpretable Model-Agnostic Explanations bar plots to interpret the model and generated directed acyclic graphs to explore the potential causal relationships between features. Additionally, we developed a web-based prediction tool using the top 10 features for AKD prediction and the top 5 features for CKD prediction. RESULTS The study cohort comprised 1559 patients. Incidence rates for AKI, AKD, and CKD were 21.7% (n=330), 15.3% (n=238), and 10.6% (n=165), respectively. Among the evaluated ML models, the Light Gradient-Boosting Machine (LightGBM) model demonstrated superior performance, with an area under the receiver operating characteristic curve of 0.97 for AKD prediction and 0.96 for CKD prediction. Performance metrics and plots highlighted the model’s competence in discrimination, calibration, and clinical applicability. Operative duration, hemoglobin, blood loss, urine protein, and hematocrit were identified as the top 5 features associated with predicted AKD. Baseline estimated glomerular filtration rate, pathology, trajectories of renal function, age, and total bilirubin were the top 5 features associated with predicted CKD. Additionally, we developed a web application using the LightGBM model to estimate AKD and CKD risks. CONCLUSIONS An interpretable ML model effectively elucidated its decision-making process in identifying patients at risk of AKD and CKD following nephrectomy by enumerating critical features. The web-based calculator, found on the LightGBM model, can assist in formulating more personalized and evidence-based clinical strategies.

Magnesium and its alloys have become potential implant materials in the future because of light weight, mechanical properties similar to natural bone, good biocompatibility, and degradability in physiological environment. However, due to the rapid corrosion and degradation of magnesium alloys in vivo, especially in the environment containing chloride ions, the application of magnesium alloys as implant materials has been limited. Therefore, improving the corrosion resistance of magnesium alloy and ensuring good biocompatibility is the main focus of the current research. In this study, hydroxyapatite coating was prepared on magnesium alloy surface using carboxymethyl cellulose-dopamine hydrogel as inducer to improve corrosion resistance and biocompatibility. Surface characterization techniques (scanning electron microscopy, Fourier-transformed infrared spectroscopy, energy dispersive X-ray spectroscopy- and X-ray diffraction) confirmed the formation of hydroxyapatite on the surface of AZ31 alloy. Corrosion resistance tests have proved the protective effect of Carboxymethyl cellulose-Dopamine/hydroxyapatite (CMC-DA/HA) coating on the surface of AZ31 alloy. According to MC3T3-E1 cell viability and Live/Dead staining, the coating also showed good biocompatibility. The results will provide new ideas for the biological application of magnesium alloys.

SummaryAccurate prediction of the solubility of chemical substances in solvents remains a challenge. The sparsity of high-quality solubility data is recognized as the biggest hurdle in the development of robust data-driven methods for practical use. Nonetheless, the effects of the quality and quantity of data on aqueous solubility predictions have not yet been scrutinized. In this study, the roles of the size and the quality of data sets on the performances of the solubility prediction models are unraveled, and the concepts of actual and observed performances are introduced. In an effort to curtail the gap between actual and observed performances, a quality-oriented data selection method, which evaluates the quality of data and extracts the most accurate part of it through statistical validation, is designed. Applying this method on the largest publicly available solubility database and using a consensus machine learning approach, a top-performing solubility prediction model is achieved.

ObjectiveThis study aimed to develop and validate interpretable machine learning (ML) models based on multi-dimensional fusion data for predicting positive surgical margins (PSM) in robot-assisted radical prostatectomy (RARP).MethodsPatients who underwent RARP at our institution between January 2016 and July 2025 were enrolled. Demographic, clinical, biopsy pathology data, and MRI-derived anatomical features (measured using ITK-SNAP on axial, sagittal, and coronal planes) were collected. Feature selection was performed using intraobserver and interobserver correlation coefficients (ICCs), low-variance filtering, univariable logistic regression, Spearman’s correlation analysis, the least absolute shrinkage and selection operator (LASSO) algorithm, and the Boruta algorithm. Six ML models were constructed, with performance evaluated using area under the curve (AUC), calibration curves, and decision curve analyses (DCA) to identify the optimal model. Five-fold and ten-fold cross-validation were used to assess the optimal model’s generalizability, and its interpretability was evaluated via Shapley Additive exPlanations (SHAP) analysis.ResultsA total of 347 patients were included, comprising a training set (n=193, January 2016–December 2024), validation set (n=84, January 2016–December 2024), and test set (n=70, January 2025–July 2025). From 164 initial features, 7 key features were retained through a four-step screening. The Random Forest (RF) model outperformed other models, achieving AUCs of 0.99 (95% CI: 0.97–1.00) in the training set, 0.88 (95% CI: 0.80–0.95) in the validation set, and 0.97 (95% CI: 0.94–1.00) in the test set. Calibration curve and decision curve analyses confirmed its strong clinical utility. Five-fold cross-validation for the RF model showed fold-specific AUCs of 0.82–0.92, with a mean AUC of 0.87 (95% CI: 0.84–0.90). Ten-fold cross-validation showed fold-specific AUCs of 0.80–0.99, with a mean AUC of 0.88 (95% CI: 0.83–0.93). SHAP analysis revealed five novel spatial anatomical features (such as Sagittal plane-posterior spatial anatomical structure index, Coronal plane-Left anatomical structure interval) were negatively associated with PSM risk, while the number of positive biopsy cores and clinical tumor stage were positively associations.ConclusionsMulti-dimensional fusion data combined with ML models improves PSM prediction accuracy in RARP. The RF model, with excellent performance and interpretability, shows promise for preoperative PSM risk stratification, facilitates optimized clinical decision-making, and supports personalized treatment discussions during preoperative planning, but requires prospective and external validation before clinical implementation.

Purpose: Lung cancer screening (LCS) lowers cancer mortality if individuals adhere to annual scans. Minoritized communities have significantly lower adherence to LCS than non-Hispanic whites. The influence of social determinants of health (SDOH) on these disparate adherence rates are rarely evaluated. The aim of this study is to develop a machine learning (ML) model that integrates individual and neighborhood level SDOH to predict which minoritized individuals are at risk of LCS non-adherence. Methods: A total of 168 participants with high-risk smoking histories who underwent their first low dose computed tomography (LDCT) between 2017 and 2021 were enrolled. Surveys on SDOH and perceived discrimination were completed. Participants’ addresses were joined to neighborhood level SDOH data from various publicly available sources. These indices captured neighborhood socioeconomic status and segregation. Adherence rates were calculated by assessing electronic medical records to determine if participants had a second LDCT within 15 months. Participants were temporally split into training (N=125) and testing (N=43) sets with a cut-off date of June 30, 2021. A gradient boosting classifier, XGBoost, was trained to predict LCS non-adherence, incorporating variables such as sociodemographic status, perceived discrimination, previous screenings, smoking status, and geospatial indices. The SHapley Additive exPlanations (SHAP) method was used to determine the relative importance of each feature in the model’s predictions. Results: The participants represented a diverse group: Asian (26%), Black/African American (32%), and Hispanic/Latino (32%). The LCS non-adherence rates were 66% in the training set and 63% in the testing set. The XGBoost classifier achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.79 and an Area Under the Precision-Recall Curve (AUPRC) of 0.90, outperforming a simple logistic regression model (AUROC=0.63 and AUPRC=0.79). The XGBoost classifier’s performance metrics were Accuracy=0.77, Recall=0.85, Specificity=0.62, Positive Predictive Value=0.79, and Negative Predictive Value=0.71. The SHAP analysis indicated that participants with no prior history of colorectal cancer screening, who lived in areas with lower quality school systems, and with a higher percentage of individuals below the poverty line, were at a higher risk of non-adherence to LCS. Conclusions: LCS reduces lung cancer survival, but LCS adherence is significantly lower in minoritized groups. Our ML model strongly predicted LCS non-adherence based on SDOH and geospatial factors. Specific neighborhood level SDOHs predicted LCS non-adherence more strongly than individual level factors. These findings support interventions that address both individual and neighborhood-level barriers to increase LCS adherence. Citation Format: Kun-Han Lu, Yi Xiao, Aamna Akhtar, Jazma Tapia, Calvin Tribby, Jay Yang, Marta Jankowska, Chi Wah Wong, Loretta Erhunmwunsee. Predicting lung cancer screening adherence among underserved and minoritized communities using social determinants and neighborhood factors [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4971.

While magnesium alloys have garnered attention for their lightweight properties across diverse applications, their susceptibility to corrosion presents a formidable challenge. Recent years have witnessed the emergence of machine learning (ML) as a formidable tool for predicting and augmenting material properties, notably corrosion resistance. This comprehensive review investigates the latest advancements and hurdles in utilizing ML techniques to investigate the corrosion behavior of magnesium alloys. This article delves into a spectrum of ML algorithms, encompassing artificial neural networks, support vector machines, and random forests, elucidating their roles in predicting corrosion rates, morphologies, and other corrosion-related characteristics in magnesium alloys. Furthermore, it underscores the pivotal challenges and opportunities within this field, such as data quality, model interpretability, and model transferability. Finally, it examines the potential of ML methods in the conception and enhancement of magnesium alloys endowed with superior corrosion resistance. This review aspires to offer valuable insights into harnessing ML’s potential for optimizing magnesium alloy designs with heightened corrosion resistance, a facet of paramount importance across diverse industries, including the automotive, aerospace, and biomedical sectors. By addressing the challenges inherent in using ML to forecast corrosion rates, including data limitations and the intricacies of corrosion mechanisms, ML stands poised to emerge as a potent instrument for advancing the development of corrosion-resistant materials.

Chronic pain around the temporomandibular joint (TMJ) and masticatory muscles is a primary symptom of temporomandibular disorders (TMD). However, the clinical significance of magnetic resonance imaging (MRI) features in predicting TMD-related pain remains unclear. This study aimed to develop and interpret machine learning (ML) models based on MRI characteristics for predicting pain severity in patients with TMD. The present retrospective study included 584 patients with TMD between January 2022 and December 2024, yielding a total of 755 TMJ MRI data sets. Pain severity was classified using the visual analogue scale (VAS). Demographic variables (age, sex) and MRI features-including lesion side, disc position, disc morphology, disc signal, disc perforation, bilaminar zone tear, joint space, joint effusion, condylar movement, bony changes and morphology/signal of the lateral pterygoid muscle-were collected. Eleven ML models based on demographic and MRI features were developed: logistic regression (LR), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), adaptive boosting (AdaBoost), gradient boosting classifier (GBC), bagging classifier (BC), extremely randomised trees (ETC), decision tree classifier (DTC) and multilayer perceptron (MLP). Model performance was evaluated using multiple metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity and F1 score. Precision-recall (PR) curves and calibration curves were plotted to assess discrimination and model calibration. Decision curve analysis (DCA) was conducted to evaluate the clinical net benefit across a range of threshold probabilities. Model interpretability was enhanced using Shapley Additive Explanations (SHAP), which quantified the contribution of each feature to individual predictions. Feature selection was conducted based on mean SHAP values, and separate LightGBM models were constructed using the Top 3, 5, and 9 most important features, as well as the full-feature set, for performance comparison. The data set was randomly divided into a training set (n = 604) and a test set (n = 151). Among the 11 ML models, the LightGBM model demonstrated the best predictive performance, with an AUC of 0.899, and was therefore identified as the optimal model. SHAP analysis identified age, disc position and condylar movement as the top three contributing features. Feature selection analysis indicated that selecting the top nine SHAP-ranked variables led to the highest diagnostic performance, with an AUC of 0.829. This study developed an interpretable, high-performing MRI-based ML model incorporating SHAP analysis to integrate imaging and clinical features for objective pain assessment, which may help identify high-risk TMD patients and guide personalised treatment strategies.

Background To assess the feasibility of a machine learning (ML) approach using radiomics features of perfusion defects on rest myocardial perfusion imaging (MPI) to detect the presence of hibernating myocardium. Methodology Data of patients who underwent 99mTc-sestamibi MPI and 18F-FDG PET/CT for myocardial viability assessment were retrieved. Rest MPI data were processed on ECToolbox, and polar maps were saved using theNFile PMap tool. The reference standard for defining hibernating myocardium was the presence of mismatched perfusion-metabolism defect with impaired myocardial contractility at rest. Perfusion defects on the polar maps were delineated with regions of interest (ROIs) after spatial resampling and intensity discretization. Replicable random sampling allocated 80% (257) of the perfusion defects of the patients from January 2017 to September 2022 to the training set and the remaining 20% (64) to the validation set. An independent dataset of perfusion defects from 29 consecutive patients from October 2022 to January 2023 was used as the testing set for model evaluation. One hundred ten first and second-order texture features were extracted for each ROI. After feature normalization and imputation, 14 best-ranked features were selected using a multistep feature selection process including the Logistic Regression and Fast Correlation-Based Filter. Thirteen supervised ML algorithms were trained with stratified five-fold cross-validation on the training set and validated on the validation set. The ML algorithms with a Log Loss of <0.688 and <0.672 in the cross-validation and validation steps were evaluated on the testing set. Performance matrices of the algorithms assessed included area under the curve (AUC), classification accuracy (CA), F1 score, precision, recall, and specificity. To provide transparency and interpretability, SHapley Additive exPlanations (SHAP) values were assessed and depicted as beeswarm plots. Results Two hundred thirty-nine patients (214 males; mean age 56 ± 11 years) were enrolled in the study. There were 371 perfusion defects (321 in the training and validation sets; 50 in the testing set). Based on the reference standard, 168 perfusion defects had hibernating myocardium (139 in the training and validation sets; 29 in the testing set). On cross-validation, six ML algorithms with Log Loss <0.688 had AUC >0.800. On validation, 10 ML algorithms had a Log Loss value <0.672, among which six had AUC >0.800. On model evaluation of the selected models on the unseen testing set, nine ML models had AUC >0.800 with Gradient Boosting Random Forest (xgboost) [GB RF (xgboost)] achieving the highest AUC of 0.860 and could detect the presence of hibernating myocardium in 21/29 (72.4%) perfusion defects with a precision of 87.5% (21/24), specificity 85.7% (18/21), CA 78.0% (39/50) and F1 Score 0.792. Four models depicted a clear pattern of model interpretability based on the beeswarm SHAP plots. These were GB RF (xgboost), GB (scikit-learn), GB (xgboost), and Random Forest. Conclusion Our study demonstrates the potential of ML in detecting hibernating myocardium using radiomics features extracted from perfusion defects on rest MPI images. This proof-of-concept underscores the notion that radiomics features capture nuanced information beyond what is perceptible to the human eye, offering promising avenues for improved myocardial viability assessment.

Background: Despite the development of disease-modifying therapy (DMT) for patients with transthyretin amyloid cardiomyopathy (ATTR-CM), the rates of mortality and disease progression are still high. The conventional models to predict prognosis (e.g., the Columbia score, the National Amyloidosis Centre [NAC] staging) predate DMT and offer limited prediction in the era of DMT. Aim: To develop a novel prognostication model in patients with ATTR-CM using plasma proteomics. Hypothesis: Plasma proteomics improves the prediction of death and disease progression in patients with ATTR-CM beyond the conventional models. Methods: In this prospective study, we conducted plasma proteomics profiling of 7,289 proteins in patients with ATTR-CM (n=303) at enrollment. The primary outcome was all-cause death, and the secondary was disease progression–a composite of death, heart transplant, heart failure hospitalization, and oral diuretics intensification. We randomly divided the cohort into training (2/3) and test sets (1/3). In the training set, we specified proteins to predict both outcomes using the Boruta algorithm. Using the specified proteins, we developed a random forest-based machine-learning (ML) model to predict each outcome in the training set. We compared the predictive ability of the ML model with the conventional models in the test set. We performed survival analyses between high- and low-risk groups defined by the ML models in the test set, while adjusting for the Columbia score. Results: During a median follow-up of 3.9 [1 st –3 rd quartile: 2.2–5.9] years, 69 patients (23%) died and 124 (41%) had disease progression, despite 243 (80%) receiving DMT. In the training set, 17 proteins were specified to predict both outcomes. In the test set, the area under the receiver-operating-characteristic curve (AUC) of the ML model was 0.90 (95% confidence interval 0.84–0.96) for all-cause death and 0.82 (0.73–0.91) for disease progression ( Image 1 ). Each ML model outperformed the conventional models in AUC, classification, and time-dependent AUC ( Images 1, 2 ). The high-risk group in the test set specified by the ML model had a higher event rate than the low-risk group in each outcome (all-cause death, adjusted hazard ratio [aHR] 6.6 [2.5–18.0], P =0.0002; disease progression, aHR 5.7 [2.6–12.4], P &lt;0.0001; Image 3 ). Conclusion: This study first demonstrated that plasma proteomics improves the prediction of mortality and disease progression in patients with ATTR-CM in the DMT era.

To develop a method for identifying high-risk patients among septic populations requiring mechanical ventilation, and to conduct phenotypic analysis based on this method. Data from four sources were utilized: the Medical Information Mart for Intensive Care (MIMIC-IV 2.0, MIMIC-III 1.4), the Philips eICU-Collaborative Research Database 2.0 (eICU-CRD 2.0), and the Anhui Medical University Second Affiliated Hospital dataset. The adult patients in intensive care unit (ICU) who met Sepsis-3 and received invasive mechanical ventilation (IMV) on the first day of first admission were enrolled. The MIMIC-IV dataset with the highest data integrity was divided into a training set and a test set at a 6:1 ratio, while the remaining datasets were served as validation sets. The demographic information, comorbidities, laboratory indicators, commonly used ICU scores, and treatment measures of patients were extracted. Clinical data collected within first day of ICU admission were used to calculate the sequential organ failure assessment (SOFA) score. K-means clustering was applied to cluster SOFA score components, and the sum of squared errors (SSE) and Davies-Bouldin index (DBI) were used to determine the optimal number of disease subtypes. For clustering results, normalized methods were employed to compare baseline characteristics by visualization, and Kaplan-Meier curves were used to analyze clinical outcomes across phenotypes. This study enrolled patients from MIMIC-IV dataset (n = 11 166), MIMIC-III dataset (n = 4 821), eICU-CRD dataset (n = 6 624), and a local dataset (n = 110), with the four datasets showing similar median ages and male proportions exceeding 50%; using 85% of the MIMIC-IV dataset as the training set, 15% as the test set, and the rest dataset as the validation set. K-means clustering based on the six-item SOFA score was performed to determine the optimal number of clusters as 3, and patients were finally classified into three phenotypes. In the training set, compared with the patients with phenotype II and phenotype III, those with phenotype I had the more severe circulatory and respiratory dysfunction, a higher proportion of vasoactive drug usage, more obvious metabolic acidosis and hypoxia, and a higher incidence of congestive heart failure. The patients with phenotype II was dominated by respiratory dysfunction with higher visceral injury. The patients with phenotype III had relatively stable organ function. The above characteristics were consistent in both the test and validation sets. Analysis of infection-related indicators showed that the patients with phenotype I had the highest SOFA score within 7 days after ICU admission, initial decreases and later increases in platelet count (PLT), and higher counts of neutrophils, lymphocytes, and monocytes as compared with those with phenotype II and phenotype III, their blood cultures had a higher positivity rates for Gram-positive bacteria, Gram-negative bacteria and fungi as compared with those with phenotype II and phenotype III. The Kaplan-Meier curve indicated that in the training, test, and validation sets, the 28-day cumulative mortality of patients with phenotype I was significantly higher than that of patients with phenotypes II and phenotype III. Three distinct phenotypes in septic patients receiving IMV based on unsupervised machine learning is derived, among which phenotype I, characterized by cardiorespiratory failure, can be used for the early identification of high-risk patients in this population. Moreover, this population is more prone to bloodstream infections, posing a high risk and having a poor prognosis.

ObjectivePostoperative complications remain a major concern in laryngeal cancer surgery, often requiring invasive interventions or intensive care. This study aimed to develop and validate an interpretable machine learning (ML) model to preoperatively predict Clavien-Dindo Grade ≥ III complications and support risk-informed perioperative decision-making.MethodsWe conducted a retrospective study using a temporally split cohort of laryngeal cancer patients. Postoperative complications were graded using the Clavien-Dindo (CD) classification. Eight ML algorithms were trained and evaluated using receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA). Model interpretability was assessed using SHapley Additive exPlanations (SHAP). A web-based calculator was deployed for clinical use.ResultThe random forest (RF) model achieved the best performance, with an area under the curve (AUC) of 0.935 in the training set and 0.842 in the test set. The model demonstrated robust sensitivity and specificity for both surgical and medical complications. Calibration curves indicated strong agreement between predicted and actual outcomes. SHAP analysis identified eight key predictors—such as vocal cord mobility, tumor subsite, and nutritional status—that contributed most to risk estimation. A user-friendly web calculator was developed and is accessible at: https://qilushiny.shinyapps.io/qilupredicate/.ConclusionWe developed a clinically interpretable ML model that accurately predicts major postoperative complications in patients undergoing laryngeal cancer surgery. This tool provides individualized risk assessments that can guide surgical planning, optimize perioperative strategies, and enhance shared decision-making. Prospective multicenter validation is needed to confirm its utility in routine practice.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12893-025-03213-z.

Field data-based prediction of local scour depth around bridge piers using interpretable machine learning

Classification of the optical spectra of active galactic nuclei (AGN) into different types is currently based on features such as line widths and intensity ratios. Although well founded on AGN physics, this approach involves some degree of human oversight and cannot scale to large datasets. Machine learning (ML) tackles this classification problem in a fast and reproducible way, but is often (and not without reason) perceived as a black box. However, ML interpretability and are active research areas in computer science that are providing us with tools to mitigate this issue. We apply ML interpretability tools to a classifier trained to predict AGN types from spectra. Our goal is to demonstrate the use of such tools in this context, obtaining for the first time insight into an otherwise black box AGN classifier. In particular, we want to understand which parts of each spectrum most affect the predictions of our classifier, checking that the results make sense in the light of our theoretical expectations. We trained a support-vector machine on 3346 high-quality, low-redshift AGN spectra from SDSS DR15. We considered either two-class classification (type 1 versus 2) or multiclass (type 1 versus 2 versus intermediate-type). The spectra were previously and independently hand-labeled and divided into types 1 and 2, and intermediate-type (i.e., sources in which the Balmer line profile consists of a sharp narrow component superimposed on a broad component). We performed a train-validation-test split, tuning hyperparameters and independently measuring performance via a variety of metrics. On a selection of test-set spectra, we computed the gradient of the predicted class probability at a given spectrum. Regions of the spectrum were then color-coded based on the direction and the amount by which they influence the predicted class, effectively building a saliency map. We also visualized the high-dimensional space of AGN spectra using t-distributed stochastic neighbor embedding (t-SNE), showing where the spectra for which we computed a saliency map are located. Our best classifier reaches an F-score of 0.942 on our test set (with 0.948 precision and 0.936 recall). We computed saliency maps on all misclassified spectra in the test set and on a sample of randomly selected spectra. Regions that affect the predicted AGN type often coincide with physically relevant features, such as spectral lines. t-SNE visualization shows good separability of type 1 and type 2 spectra. Intermediate-type spectra either lie in-between, as expected, or appear mixed with type 2 spectra. Misclassified spectra are typically found among the latter. Some clustering structure is apparent among type 2 and intermediate-type spectra, though this may be an artifact. Saliency maps show why a given AGN type was predicted by our classifier resulting in a physical interpretation in terms of regions of the spectrum that affected its decision, making it no longer a black box. These regions coincide with those used by human experts, for example relevant spectral lines, and are even used in a similar way; the classifier effectively measures the width of a line by weighing its center and its tails oppositely.