This scoping review focuses on the essential role of models for causal inference in shaping actionable artificial intelligence (AI) designed to aid clinicians in decision-making. The objective was to identify and evaluate the reporting quality of studies introducing models for causal inference in intensive care units (ICUs), and to provide recommendations to improve the future landscape of research practices in this domain. To achieve this, we searched various databases including Embase, MEDLINE ALL, Web of Science Core Collection, Google Scholar, medRxiv, bioRxiv, arXiv, and the ACM Digital Library. Studies involving models for causal inference addressing time-varying treatments in the adult ICU were reviewed. Data extraction encompassed the study settings and methodologies applied. Furthermore, we assessed reporting quality of target trial components (i.e., eligibility criteria, treatment strategies, follow-up period, outcome, and analysis plan) and main causal assumptions (i.e., conditional exchangeability, positivity, and consistency). Among the 2184 titles screened, 79 studies met the inclusion criteria. The methodologies used were G methods (61%) and reinforcement learning methods (39%). Studies considered both static (51%) and dynamic treatment regimes (49%). Only 30 (38%) of the studies reported all five target trial components, and only seven (9%) studies mentioned all three causal assumptions. To achieve actionable AI in the ICU, we advocate careful consideration of the causal question of interest, describing this research question as a target trial emulation, usage of appropriate causal inference methods, and acknowledgement (and examination of potential violations of) the causal assumptions.

@en

Artificial intelligence (AI) research in the intensive care unit (ICU) mainly focuses on developing models (from linear regression to deep learning) to predict out-
comes, such as mortality or sepsis [1, 2]. However, there is another important aspect of AI that is typically not framed as AI (although it may be more worthy of the name), which is the prediction of patient outcomes or events that would result from different actions, known as causal inference [3, 4]. This aspect of AI is crucial for decision-making in the ICU. To emphasize the impor- tance of causal inference, we propose to refer to any data- driven model used for causal inference tasks as ‘action- able AI’, as opposed to ‘predictive AI’, and discuss how these models could provide meaningful decision support in the ICU.@en

We study the problem of falsifying the assumptions behind a set of broadly applied causal identification strategies: namely back-door adjustment, front-door adjustment, and instrumental variable estimation. While these assumptions are untestable from observational data in general, we show that with access to data coming from multiple heterogeneous environments, there exist novel independence constraints that can be used to falsify the validity of each strategy. Most interestingly, we make no parametric assumptions, instead relying on that changes between environments happen under the principle of independent causal mechanisms. @en

A common assumption in causal inference from observational data is that there is no hidden confounding. Yet it is, in general, impossible to verify this assumption from a single dataset. Under the assumption of independent causal mechanisms underlying the data-generating process, we demonstrate a way to detect unobserved confounders when having multiple observational datasets coming from different environments. We present a theory for testable conditional independencies that are only absent when there is hidden confounding and examine cases where we violate its assumptions: degenerate & dependent mechanisms, and faithfulness violations. Additionally, we propose a procedure to test these independencies and study its empirical finite-sample behavior using simulation studies and semi-synthetic data based on a real-world dataset. In most cases, the proposed procedure correctly predicts the presence of hidden confounding, particularly when the confounding bias is large. @en

Arguably, a desirable feature of a learner is that its performance gets better with an increasing amount of training data, at least in expectation. This issue has received renewed attention in recent years and some curious and surprising findings have been reported on. In essence, these results show that more data does actually not necessarily lead to improved performance—worse even, performance can deteriorate. Clustering, however, has not been subjected to such kind of study up to now. This paper shows that k-means clustering, a ubiquitous technique in machine learning and data mining, suffers from the same lack of so-called monotonicity and can display deterioration in expected performance with increasing training set sizes. Our main, theoretical contributions prove that 1-means clustering is monotonic, while 2-means is not even weakly monotonic, i.e., the occurrence of nonmonotonic behavior persists indefinitely, beyond any training sample size. For larger k, the question remains open.

@en

Background: Currently available treatment options for Parkinson's disease are symptomatic and do not alter the course of the disease. Recent studies have raised the possibility that cardiovascular risk management may slow the progression of the disease. Objectives: We estimated the effect of baseline cardiovascular risk factors on the progression of Parkinson's disease, using measures for PD-specific motor signs and cognitive functions. Methods: We used data from 424 de novo Parkinson's disease patients and 199 age-matched controls from the observational, multicenter Parkinson's Progression Markers Initiative (PPMI) study, which included follow-up of up to 9 years. The primary outcome was the severity of PD-specific motor signs, assessed with the MDS-UPDRS part III in the “OFF”-state. The secondary outcome was cognitive function, measured with the Montreal Cognitive Assessment, Symbol Digit Modalities Test, and Letter-Number Sequencing task. Exposures of interest were diabetes mellitus, hypertension, body mass index, cardiovascular event history and hypercholesterolemia, and a modified Framingham risk score, measured at baseline. The effect of each of these exposures on disease progression was modeled using linear mixed models, including adjustment for identified confounders. A secondary analysis on the Tracking Parkinson's cohort including 1,841 patients was performed to validate our findings in an independent patient cohort. Results: Mean age was 61.4 years, and the average follow-up was 5.5 years. We found no statistically significant effect of any individual cardiovascular risk factor on the MDS-UPDRS part III progression (all 95% confidence intervals (CIs) included zero), with one exception: in the PD group, the estimated effect of a one-point increase in body mass index was 0.059 points on the MDS-UPDRS part III per year (95% CI: 0.017 to 0.102). We found no evidence for an effect of any of the exposures on the rate of change in cognitive functioning in the PD group. Similar results were observed for the Tracking Parkinson's cohort (all 95% CIs overlapped with PPMI), but the 95% CI of the effect of body mass index on the MDS-UPDRS part III progression included zero. Conclusions: Based on this analysis of two large cohorts of de novo PD patients, we found no evidence to support clinically relevant effects of cardiovascular risk factors on the clinical progression of Parkinson's disease.

@en

Learning curves illustrate how generalization performance of a learner evolves with more training data. While this is a useful tool to characterize learners, not all learning curve behavior is well understood. For instance, it is sometimes assumed that the more training data provided, the better the learner performs. However, counter-examples exist for both classical machine learning algorithms and deep neural networks, where errors do not monotonically decrease with training set size. Loog et al. [12] describe this monotonicity problem, and present several regression examples where simple empirical risk minimizers display unexpected learning curve behaviors. In this paper, we will study two of these proposed problems in detail and explain what caused the odd learning curves. For the first, we use a bias-variance decomposition to show that the monotonic increase in the learning curve is caused by an increase in the variance, which we explain by a mismatch between the model and the data generating process. For the second problem, we explain the recurring increases in the learning curve by showing only two solutions are attainable by the learner. The probability of obtaining a configuration of training objects that leads to the high risk solution typically decreases as the training set size increases. However, for particular training set sizes, additional configurations that produce the high risk solution become possible. We prove that these additional configurations increase the probability of the high risk solution and therefore explain the unusual learning curve. These examples contribute to a more complete understanding of learning curves and the possibilities and reasons behind their various behaviors.

@en

Background: Optimal care for Parkinson’s disease (PD) requires coordination and collaboration between providers within a complex care network. Individual patients have personalised networks of their own providers, creating a unique informal network of providers who treat (‘share’) the same patient. These ‘patient-sharing networks’ differ in density, ie, the number of identical patients they share. Denser patient-sharing networks might reflect better care provision, since providers who share many patients might have made efforts to improve their mutual care delivery. We evaluated whether the density of these patient-sharing networks affects patient outcomes and costs. Methods: We analysed medical claims data from all PD patients in the Netherlands between 2012 and 2016. We focused on seven professional disciplines that are commonly involved in Parkinson care. We calculated for each patient the density score: the average number of patients that each patient’s providers shared. Density scores could range from 1.00 (which might reflect poor collaboration) to 83.00 (which might reflect better collaboration). This score was also calculated at the hospital level by averaging the scores for all patients belonging to a specific hospital. Using logistic and linear regression analyses we estimated the relationship between density scores and health outcomes, healthcare utilization, and healthcare costs. Results: The average density score varied considerably (average 6.7, SD 8.2). Adjusted for confounders, higher density scores were associated with a lower risk of PD-related complications (odds ratio [OR]: 0.901; P <.001) and with lower healthcare costs (coefficients:-0.018, P =.005). Higher density scores were associated with more frequent involvement of neurologists (coefficient 0.068), physiotherapists (coefficient 0.052) and occupational therapists (coefficient 0.048) (P values all <.001). Conclusion: Patient sharing networks showed large variations in density, which appears unwanted as denser networks are associated with better outcomes and lower costs.

@en

Introduction: Unplanned hospital admissions associated with Parkinson's disease could be partly attributable to comorbidities. Methods: We studied nationwide claims databases and registries. Persons with newly diagnosed Parkinson's disease were identified based on the first Parkinson's disease-related reimbursement claim by a medical specialist. Comorbidities were classified based on the Charlson Comorbidity Index. We studied hospitalization admissions because of falls, psychiatric diseases, pneumonia and urinary tract infections, PD-related hospitalizations-not otherwise specified. The association between comorbidities and time-to-hospitalization was estimated using Cox proportional hazard modelling. To better understand pathways leading to hospitalizations, we performed multiple analyses on causes for hospitalizations. Results: We identified 18 586 people with newly diagnosed Parkinson's disease. The hazard of hospitalization was increased in persons with peptic ulcer disease (HR 2.20, p = 0.009), chronic obstructive pulmonary disease (HR 1.61, p < 0.001), stroke (HR 1.37, p = 0.002) and peripheral vascular disease (HR 1.31, p = 0.02). In the secondary analyses, the hazard of PD-related hospitalizations-not otherwise specified (HR 3.24, p = 0.02) and pneumonia-related hospitalization (HR 2.90, p = 0.03) was increased for those with comorbid peptic ulcer disease. The hazard of fall-related hospitalization (HR 1.57, p = 0.003) and pneumonia-related hospitalization (HR 2.91, p < 0.001) was increased in persons with chronic obstructive pulmonary disease. The hazard of pneumonia-related hospitalization was increased in those with stroke (HR 1.54, p = 0.03) or peripheral vascular disease (HR 1.60, p = 0.02). The population attributable risk of comorbidity was 8.4%. Conclusion: Several comorbidities increase the risk of Parkinson's disease related-hospitalization indicating a need for intervention strategies targeting these comorbid disorders.

@en

Background: The COVID-19 pandemic continues to overwhelm intensive care units (ICUs) worldwide, and improved prediction of mortality among COVID-19 patients could assist decision making in the ICU setting. In this work, we report on the development and validation of a dynamic mortality model specifically for critically ill COVID-19 patients and discuss its potential utility in the ICU. Methods: We collected electronic medical record (EMR) data from 3222 ICU admissions with a COVID-19 infection from 25 different ICUs in the Netherlands. We extracted daily observations of each patient and fitted both a linear (logistic regression) and non-linear (random forest) model to predict mortality within 24 h from the moment of prediction. Isotonic regression was used to re-calibrate the predictions of the fitted models. We evaluated the models in a leave-one-ICU-out (LOIO) cross-validation procedure. Results: The logistic regression and random forest model yielded an area under the receiver operating characteristic curve of 0.87 [0.85; 0.88] and 0.86 [0.84; 0.88], respectively. The recalibrated model predictions showed a calibration intercept of −0.04 [−0.12; 0.04] and slope of 0.90 [0.85; 0.95] for logistic regression model and a calibration intercept of −0.19 [−0.27; −0.10] and slope of 0.89 [0.84; 0.94] for the random forest model. Discussion: We presented a model for dynamic mortality prediction, specifically for critically ill COVID-19 patients, which predicts near-term mortality rather than in-ICU mortality. The potential clinical utility of dynamic mortality models such as benchmarking, improving resource allocation and informing family members, as well as the development of models with more causal structure, should be topics for future research.

@en

Background: Timely identification of deteriorating COVID-19 patients is needed to guide changes in clinical management and admission to intensive care units (ICUs). There is significant concern that widely used Early warning scores (EWSs) underestimate illness severity in COVID-19 patients and therefore, we developed an early warning model specifically for COVID-19 patients. Methods: We retrospectively collected electronic medical record data to extract predictors and used these to fit a random forest model. To simulate the situation in which the model would have been developed after the first and implemented during the second COVID-19 ‘wave’ in the Netherlands, we performed a temporal validation by splitting all included patients into groups admitted before and after August 1, 2020. Furthermore, we propose a method for dynamic model updating to retain model performance over time. We evaluated model discrimination and calibration, performed a decision curve analysis, and quantified the importance of predictors using SHapley Additive exPlanations values. Results: We included 3514 COVID-19 patient admissions from six Dutch hospitals between February 2020 and May 2021, and included a total of 18 predictors for model fitting. The model showed a higher discriminative performance in terms of partial area under the receiver operating characteristic curve (0.82 [0.80–0.84]) compared to the National early warning score (0.72 [0.69–0.74]) and the Modified early warning score (0.67 [0.65–0.69]), a greater net benefit over a range of clinically relevant model thresholds, and relatively good calibration (intercept = 0.03 [− 0.09 to 0.14], slope = 0.79 [0.73–0.86]). Conclusions: This study shows the potential benefit of moving from early warning models for the general inpatient population to models for specific patient groups. Further (independent) validation of the model is needed.

@en

We present ReproducedPapers.org : an open online repository for teaching and structuring machine learning reproducibility. We evaluate doing a reproduction project among students and the added value of an online reproduction repository among AI researchers. We use anonymous self-assessment surveys and obtained 144 responses. Results suggest that students who do a reproduction project place more value on scientific reproductions and become more critical thinkers. Students and AI researchers agree that our online reproduction repository is valuable.

@en

Background: Female childhood cancer survivors (CCSs) carry a risk of therapy-related gonadal dysfunction. Alkylating agents (AA) are well-established risk factors, yet inter-individual variability in ovarian function is observed. Polymorphisms in CYP450 enzymes may explain this variability in AA-induced ovarian damage. We aimed to evaluate associations between previously identified genetic polymorphisms in CYP450 enzymes and AA-related ovarian function among adult CCSs. Methods: Anti-Müllerian hormone (AMH) levels served as a proxy for ovarian function in a discovery cohort of adult female CCSs, from the pan-European PanCareLIFE cohort (n = 743; age (years): Median 25.8, interquartile range (IQR) 22.1-30.6). Using two additive genetic models in linear and logistic regression, nine genetic variants in three CYP450 enzymes were analyzed in relation to cyclophosphamide equivalent dose (CED) score and their impact on AMH levels. The main model evaluated the effect of the variant on AMH and the interaction model evaluated the modifying effect of the variant on the impact of CED score on log-transformed AMH levels. Results were validated, and meta-analysis performed, using the USA-based St. Jude Lifetime Cohort (n = 391; age (years): Median 31.3, IQR 26.6-37.4). Results: CYP3A4*3 was significantly associated with AMH levels in the discovery and replication cohort. Meta-analysis revealed a significant main deleterious effect (Beta (95% CI): −0.706 (−1.11-−0.298), p-value = 7 × 10−4) of CYP3A4*3 (rs4986910) on log-transformed AMH levels. CYP2B6*2 (rs8192709) showed a significant protective interaction effect (Beta (95% CI): 0.527 (0.126-0.928), p-value = 0.01) on log-transformed AMH levels in CCSs receiving more than 8000 mg/m² CED. Conclusions: Female CCSs CYP3A4*3 carriers had significantly lower AMH levels, and CYP2B6*2 may have a protective effect on AMH levels. Identification of risk-contributing variants may improve individualized counselling regarding the treatment-related risk of infertility and fertility preservation options.

@en

STUDY QUESTION: Do genetic variations in the DNA damage response pathway modify the adverse effect of alkylating agents on ovarian function in female childhood cancer survivors (CCS)? SUMMARY ANSWER: Female CCS carrying a common BR serine/threonine kinase 1 (BRSK1) gene variant appear to be at 2.5-fold increased odds of reduced ovarian function after treatment with high doses of alkylating chemotherapy. WHAT IS KNOWN ALREADY: Female CCS show large inter-individual variability in the impact of DNA-damaging alkylating chemotherapy, given as treatment of childhood cancer, on adult ovarian function. Genetic variants in DNA repair genes affecting ovarian function might explain this variability. STUDY DESIGN, SIZE, DURATION: CCS for the discovery cohort were identified from the Dutch Childhood Oncology Group (DCOG) LATER VEVO-study, a multi-centre retrospective cohort study evaluating fertility, ovarian reserve and risk of premature menopause among adult female 5-year survivors of childhood cancer. Female 5-year CCS, diagnosed with cancer and treated with chemotherapy before the age of 25 years, and aged 18 years or older at time of study were enrolled in the current study. Results from the discovery Dutch DCOG-LATER VEVO cohort (n = 285) were validated in the pan-European PanCareLIFE (n = 465) and the USA-based St. Jude Lifetime Cohort (n = 391). PARTICIPANTS/MATERIALS, SETTING, METHODS: To evaluate ovarian function, anti-Müllerian hormone (AMH) levels were assessed in both the discovery cohort and the replication cohorts. Using additive genetic models in linear and logistic regression, five genetic variants involved in DNA damage response were analysed in relation to cyclophosphamide equivalent dose (CED) score and their impact on ovarian function. Results were then examined using fixed-effect meta-analysis. MAIN RESULTS AND THE ROLE OF CHANCE: Meta-analysis across the three independent cohorts showed a significant interaction effect (P = 3.0 × 10-4) between rs11668344 of BRSK1 (allele frequency = 0.34) among CCS treated with high-dose alkylating agents (CED score ≥8000 mg/m2), resulting in a 2.5-fold increased odds of a reduced ovarian function (lowest AMH tertile) for CCS carrying one G allele compared to CCS without this allele (odds ratio genotype AA: 2.01 vs AG: 5.00). LIMITATIONS, REASONS FOR CAUTION: While low AMH levels can also identify poor responders in assisted reproductive technology, it needs to be emphasized that AMH remains a surrogate marker of ovarian function. WIDER IMPLICATIONS OF THE FINDINGS: Further research, validating our findings and identifying additional risk-contributing genetic variants, may enable individualized counselling regarding treatment-related risks and necessity of fertility preservation procedures in girls with cancer. STUDY FUNDING/COMPETING INTEREST(S): This work was supported by the PanCareLIFE project that has received funding from the European Union's Seventh Framework Programme for research, technological development and demonstration under grant agreement no 602030. In addition, the DCOG-LATER VEVO study was funded by the Dutch Cancer Society (Grant no. VU 2006-3622) and by the Children Cancer Free Foundation (Project no. 20) and the St Jude Lifetime cohort study by NCI U01 CA195547. The authors declare no competing interests. TRIAL REGISTRATION NUMBER: N/A.

@en

Veel patiënten met een complexe neurologische aandoening, zoals de ziekte van Parkinson, multiple sclerose of restverschijnselen van niet-aangeboren hersenletsel, doen vroeg of laat in hun ziekteproces een beroep op de langdurige zorg. Dat deze mensen een andere zorgbehoefte hebben dan de algemene verpleeghuispopulatie is bekend, maar van hun aantallen bestaan slechts schattingen. Dit onderzoek, mogelijk gemaakt door de Verenso Beurs voor wetenschappelijk onderzoek, heeft als doel te achterhalen hoeveel mensen in Nederland een indicatie voor langdurige zorg hebben én een diagnose die wijst op een ernstige aandoening van het zenuwstelsel, anders dan dementie. We combineerden grote datasets van het Centraal Bureau voor Statistiek om zo de prevalentie van een aantal complexe neurologische aandoeningen onder mensen met een indicatie voor langdurige zorg in een verpleeghuis te bepalen. Op 31-12-2015 hadden 121.749 mensen in Nederland een indicatie die toegang geeft tot langdurige zorg in een verpleeghuis. Van hen hebben 9.398 mensen een complexe neurologische aandoening. De ziekte van Parkinson en aanverwante aandoeningen komen bij 4% van de totale langdurige-zorgpopulatie voor. Patiënten met een zeldzamere aandoening zijn 10 tot 20 jaar jonger dan de gemiddelde verpleeghuispatiënt. Er zijn in de langdurige zorg meer mensen met een complexe neurologische aandoening, anders dan dementie, dan tot nu toe werd aangenomen. Door verdere academisering van de langdurige zorg en de ontwikkeling van doelgroepexpertisecentra moet het mogelijk zijn ook deze patiënten en hun naasten passende zorg te bieden. @en

Background: Both patients and physicians may choose to delay initiation of dopamine replacement therapy in Parkinson's disease (PD) for various reasons. We used observational data to estimate the effect of earlier treatment in PD. Observational data offer a valuable source of evidence, complementary to controlled trials. Method: We studied the Parkinson's Progression Markers Initiative cohort of patients with de novo PD to estimate the effects of duration of PD treatment during the first 2 years of follow-up, exploiting natural interindividual variation in the time to start first treatment. We estimated the Movement Disorder Society–Unified Parkinson's Disease Rating Scale (MDS-UPDRS) Part III (primary outcome) and several functionally relevant outcomes at 2, 3, and 4 years after baseline. To adjust for time-varying confounding, we used marginal structural models with inverse probability of treatment weighting and the parametric g-formula. Results: We included 302 patients from the Parkinson's Progression Markers Initiative cohort. There was a small improvement in MDS-UPDRS Part III scores after 2 years of follow-up for patients who started treatment earlier, and similar, but nonstatistically significant, differences in subsequent years. We found no statistically significant differences in most secondary outcomes, including the presence of motor fluctuations, nonmotor symptoms, MDS-UPDRS Part II scores, and the Schwab and England Activities of Daily Living Scale. Conclusion: Earlier treatment initiation does not lead to worse MDS-UPDRS motor scores and may offer small improvements. These findings, based on observational data, are in line with earlier findings from clinical trials. Observational data, when combined with appropriate causal methods, are a valuable source of additional evidence to support real-world clinical decisions.

@en

Background: Wearable sensors have been used successfully to characterize bradykinetic gait in patients with Parkinson disease (PD), but most studies to date have been conducted in highly controlled laboratory environments. Objective: This paper aims to assess whether sensor-based analysis of real-life gait can be used to objectively and remotely monitor motor fluctuations in PD. Methods: The Parkinson@Home validation study provides a new reference data set for the development of digital biomarkers to monitor persons with PD in daily life. Specifically, a group of 25 patients with PD with motor fluctuations and 25 age-matched controls performed unscripted daily activities in and around their homes for at least one hour while being recorded on video. Patients with PD did this twice: once after overnight withdrawal of dopaminergic medication and again 1 hour after medication intake. Participants wore sensors on both wrists and ankles, on the lower back, and in the front pants pocket, capturing movement and contextual data. Gait segments of 25 seconds were extracted from accelerometer signals based on manual video annotations. The power spectral density of each segment and device was estimated using Welch's method, from which the total power in the 0.5- to 10-Hz band, width of the dominant frequency, and cadence were derived. The ability to discriminate between before and after medication intake and between patients with PD and controls was evaluated using leave-one-subject-out nested cross-validation. Results: From 18 patients with PD (11 men; median age 65 years) and 24 controls (13 men; median age 68 years), ≥10 gait segments were available. Using logistic LASSO (least absolute shrinkage and selection operator) regression, we classified whether the unscripted gait segments occurred before or after medication intake, with mean area under the receiver operator curves (AUCs) varying between 0.70 (ankle of least affected side, 95% CI 0.60-0.81) and 0.82 (ankle of most affected side, 95% CI 0.72-0.92) across sensor locations. Combining all sensor locations did not significantly improve classification (AUC 0.84, 95% CI 0.75-0.93). Of all signal properties, the total power in the 0.5- to 10-Hz band was most responsive to dopaminergic medication. Discriminating between patients with PD and controls was generally more difficult (AUC of all sensor locations combined: 0.76, 95% CI 0.62-0.90). The video recordings revealed that the positioning of the hands during real-life gait had a substantial impact on the power spectral density of both the wrist and pants pocket sensor. Conclusions: We present a new video-referenced data set that includes unscripted activities in and around the participants' homes. Using this data set, we show the feasibility of using sensor-based analysis of real-life gait to monitor motor fluctuations with a single sensor location. Future work may assess the value of contextual sensors to control for real-world confounders.

@en

Active learning algorithms propose what data should be labeled given a pool of unlabeled data. Instead of selecting randomly what data to annotate, active learning strategies aim to select data so as to get a good predictive model with as little labeled samples as possible. Single-shot batch active learners select all samples to be labeled in a single step, before any labels are observed.We study single-shot active learners that minimize generalization bounds to select a representative sample, such as the maximum mean discrepancy (MMD) active learner.We prove that a related bound, the discrepancy, provides a tighter worst-case bound. We study these bounds probabilistically, which inspires us to introduce a novel bound, the nuclear discrepancy (ND). The ND bound is tighter for the expected loss under optimistic probabilistic assumptions. Our experiments show that the MMD active learner performs better than the discrepancy in terms of the mean squared error, indicating that tighter worst case bounds do not imply better active learning performance. The proposed active learner improves significantly upon the MMD and discrepancy in the realizable setting and a similar trend is observed in the agnostic setting, showing the benefits of a probabilistic approach to active learning. Our study highlights that assumptions underlying generalization bounds can be equally important as bound-tightness, when it comes to active learning performance. Code for reproducing our experimental results can be found at https://github.com/tomviering/ NuclearDiscrepancy. @en