This file was created by the TYPO3 extension bib --- Timezone: CEST Creation date: 2024-04-26 Creation time: 01-42-33 --- Number of references 22 article 10.3389/frai.2024.1330919 Frontiers in Artificial Intelligence 2024 7 Convolutional Neural Networks (CNNs) are frequently and successfully used in medical prediction tasks. They are often used in combination with transfer learning, leading to improved performance when training data for the task are scarce. The resulting models are highly complex and typically do not provide any insight into their predictive mechanisms, motivating the field of “explainable” artificial intelligence (XAI). However, previous studies have rarely quantitatively evaluated the “explanation performance” of XAI methods against ground-truth data, and transfer learning and its influence on objective measures of explanation performance has not been investigated. Here, we propose a benchmark dataset that allows for quantifying explanation performance in a realistic magnetic resonance imaging (MRI) classification task. We employ this benchmark to understand the influence of transfer learning on the quality of explanations. Experimental results show that popular XAI methods applied to the same underlying model differ vastly in performance, even when considering only correctly classified examples. We further observe that explanation performance strongly depends on the task used for pre-training and the number of CNN layers pre-trained. These results hold after correcting for a substantial correlation between explanation and classification performance. 8.4,8.44 https://www.frontiersin.org/articles/10.3389/frai.2024.1330919 2624-8212 10.3389/frai.2024.1330919 MartaOliveira RickWilming BenedictClark CélineBudding FabianEitel KerstinRitter StefanHaufe inproceedings pmlr-v202-wilming23a 2023 7 202 37091--37107 In recent years, the community of ’explainable artificial intelligence’ (XAI) has created a vast body of methods to bridge a perceived gap between model ’complexity’ and ’interpretability’. However, a concrete problem to be solved by XAI methods has not yet been formally stated. As a result, XAI methods are lacking theoretical and empirical evidence for the ’correctness’ of their explanations, limiting their potential use for quality-control and transparency purposes. At the same time, Haufe et al. (2014) showed, using simple toy examples, that even standard interpretations of linear models can be highly misleading. Specifically, high importance may be attributed to so-called suppressor variables lacking any statistical relation to the prediction target. This behavior has been confirmed empirically for a large array of XAI methods in Wilming et al. (2022). Here, we go one step further by deriving analytical expressions for the behavior of a variety of popular XAI methods on a simple two-dimensional binary classification problem involving Gaussian class-conditional distributions. We show that the majority of the studied approaches will attribute non-zero importance to a non-class-related suppressor feature in the presence of correlated noise. This poses important limitations on the interpretations and conclusions that the outputs of these XAI methods can afford. 8.4,8.44 https://proceedings.mlr.press/v202/wilming23a.html Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan PMLR Proceedings of Machine Learning Research Proceedings of the 40th International Conference on Machine Learning RickWilming LeoKieslich BenedictClark StefanHaufe article cesnaite_alterations_2023 NeuroImage 2023 268 119810 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1053811922009314 2023-08-22 en 10538119 10.1016/j.neuroimage.2022.119810 ElenaCesnaite PaulSteinfath MinaJamshidi Idaji TilmanStephani DenizKumral StefanHaufe ChristianSander TilmanHensch UlrichHegerl SteffiRiedel-Heller SusanneRöhr Matthias L.Schroeter A．VeronicaWitte ArnoVillringer Vadim V.Nikulin article pellegrini_identifying_2023 NeuroImage 2023 277 120218 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1053811923003695 2023-08-22 en 10538119 10.1016/j.neuroimage.2023.120218 FranziskaPellegrini ArnaudDelorme VadimNikulin StefanHaufe article cai_bayesian_2023 IEEE Transactions on Medical Imaging 2023 42 3 762--773 8.4,8.44 https://ieeexplore.ieee.org/document/9932439/ 2023-08-22 0278-0062, 1558-254X 10.1109/TMI.2022.3218074 ChangCai HuicongKang AliHashemi DanChen MithunDiwakar StefanHaufe KensukeSekihara WeiWu Srikantan S.Nagarajan article haufe_gait_2023 Bioengineering 2023 10 2 212 Gait disturbances are common manifestations of Parkinson’s disease (PD), with unmet therapeutic needs. Inertial measurement units (IMUs) are capable of monitoring gait, but they lack neurophysiological information that may be crucial for studying gait disturbances in these patients. Here, we present a machine learning approach to approximate IMU angular velocity profiles and subsequently gait events using electromyographic (EMG) channels during overground walking in patients with PD. We recorded six parkinsonian patients while they walked for at least three minutes. Patient-agnostic regression models were trained on temporally embedded EMG time series of different combinations of up to five leg muscles bilaterally (i.e., tibialis anterior, soleus, gastrocnemius medialis, gastrocnemius lateralis, and vastus lateralis). Gait events could be detected with high temporal precision (median displacement of <prt>\textless</prt>50 ms), low numbers of missed events (<prt>\textless</prt>2%), and next to no false-positive event detections (<prt>\textless</prt>0.1%). Swing and stance phases could thus be determined with high fidelity (median F1-score of <prt>\textasciitilde</prt>0.9). Interestingly, the best performance was obtained using as few as two EMG probes placed on the left and right vastus lateralis. Our results demonstrate the practical utility of the proposed EMG-based system for gait event prediction, which allows the simultaneous acquisition of an electromyographic signal to be performed. This gait analysis approach has the potential to make additional measurement devices such as IMUs and force plates less essential, thereby reducing financial and preparation overheads and discomfort factors in gait studies. 8.4, 8.44 https://www.mdpi.com/2306-5354/10/2/212 2023-08-22 en 2306-5354 10.3390/bioengineering10020212 StefanHaufe Ioannis U.Isaias FranziskaPellegrini ChiaraPalmisano article cai_bayesian_2023-1 IEEE Transactions on Medical Imaging 2023 1--1 8.4,8.44 https://ieeexplore.ieee.org/document/10068553/ 2023-08-22 0278-0062, 1558-254X 10.1109/TMI.2023.3256963 ChangCai YuanshunLong SanjayGhosh AliHashemi YijingGao MithunDiwakar StefanHaufe KensukeSekihara WeiWu Srikantan S.Nagarajan article del_vecchio_del_vecchio_pallidal_2023 Bioengineering 2023 10 4 476 Low-frequency oscillatory patterns of pallidal local field potentials (LFPs) have been proposed as a physiomarker for dystonia and hold the promise for personalized adaptive deep brain stimulation. Head tremor, a low-frequency involuntary rhythmic movement typical of cervical dystonia, may cause movement artifacts in LFP signals, compromising the reliability of low-frequency oscillations as biomarkers for adaptive neurostimulation. We investigated chronic pallidal LFPs with the PerceptTM PC (Medtronic PLC) device in eight subjects with dystonia (five with head tremors). We applied a multiple regression approach to pallidal LFPs in patients with head tremors using kinematic information measured with an inertial measurement unit (IMU) and an electromyographic signal (EMG). With IMU regression, we found tremor contamination in all subjects, whereas EMG regression identified it in only three out of five. IMU regression was also superior to EMG regression in removing tremor-related artifacts and resulted in a significant power reduction, especially in the theta-alpha band. Pallido-muscular coherence was affected by a head tremor and disappeared after IMU regression. Our results show that the Percept PC can record low-frequency oscillations but also reveal spectral contamination due to movement artifacts. IMU regression can identify such artifact contamination and be a suitable tool for its removal. 8.4, 8.44 https://www.mdpi.com/2306-5354/10/4/476 2023-08-22 en 2306-5354 10.3390/bioengineering10040476 JasminDel Vecchio Del Vecchio IbrahemHanafi Nicoló GabrielePozzi PhilippCapetian Ioannis U.Isaias StefanHaufe ChiaraPalmisano article cai_empirical_2022 NeuroImage 2022 258 119369 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1053811922004888 2022-09-13 en 10538119 10.1016/j.neuroimage.2022.119369 ChangCai LeightonHinkley YijingGao AliHashemi StefanHaufe KensukeSekihara Srikantan S.Nagarajan article langer_benchmark_2022 NeuroImage 2022 258 119348 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1053811922004670 2022-09-13 en 10538119 10.1016/j.neuroimage.2022.119348 NicolasLanger Martyna BeataPlomecka MariusTröndle AnujaNegi TzvetanPopov MichaelMilham StefanHaufe article wilming_scrutinizing_2022 Machine Learning 2022 111 5 1903--1923 Abstract Machine learning (ML) is increasingly often used to inform high-stakes decisions. As complex ML models (e.g., deep neural networks) are often considered black boxes, a wealth of procedures has been developed to shed light on their inner workings and the ways in which their predictions come about, defining the field of ‘explainable AI’ (XAI). Saliency methods rank input features according to some measure of ‘importance’. Such methods are difficult to validate since a formal definition of feature importance is, thus far, lacking. It has been demonstrated that some saliency methods can highlight features that have no statistical association with the prediction target (suppressor variables). To avoid misinterpretations due to such behavior, we propose the actual presence of such an association as a necessary condition and objective preliminary definition for feature importance. We carefully crafted a ground-truth dataset in which all statistical dependencies are well-defined and linear, serving as a benchmark to study the problem of suppressor variables. We evaluate common explanation methods including LRP, DTD, PatternNet, PatternAttribution, LIME, Anchors, SHAP, and permutation-based methods with respect to our objective definition. We show that most of these methods are unable to distinguish important features from suppressors in this setting. 8.4, 8.44 https://link.springer.com/10.1007/s10994-022-06167-y 2022-09-13 en 0885-6125, 1573-0565 10.1007/s10994-022-06167-y RickWilming CélineBudding Klaus-RobertMüller StefanHaufe article kumral_relationship_2022 Neurobiology of Aging 2022 112 1--11 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S0197458021003195 2022-09-13 en 01974580 10.1016/j.neurobiolaging.2021.10.006 DenizKumral ElenaCesnaite FraukeBeyer Simon M.Hofmann TilmanHensch ChristianSander UlrichHegerl StefanHaufe ArnoVillringer A. VeronicaWitte Vadim V.Nikulin article merk_machine_2022 Experimental Neurology 2022 351 113993 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S0014488622000188 2022-09-13 en 00144886 10.1016/j.expneurol.2022.113993 TimonMerk VictoriaPeterson RichardKöhler StefanHaufe R. MarkRichardson Wolf-JulianNeumann article hashemi_joint_2022 IEEE Transactions on Medical Imaging 2022 1--1 8.4, 8.44 https://ieeexplore.ieee.org/document/9963991/ 2023-08-22 0278-0062, 1558-254X 10.1109/TMI.2022.3224085 AliHashemi ChangCai YijingGao SanjayGhosh Klaus-RobertMüller Srikantan S.Nagarajan StefanHaufe article pellegrini_p_2022 Clinical Neurophysiology 2022 137 e60--e61 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1388245722001298 2023-08-22 en 13882457 10.1016/j.clinph.2022.01.111 F.Pellegrini V.Nikulin S.Haufe article noauthor_erratum:_2022 The Journal of Neuroscience 2022 42 22 4605--4605 8.4, 8.44 https://www.jneurosci.org/lookup/doi/10.1523/JNEUROSCI.0820-22.2022 2023-08-22 en 0270-6474, 1529-2401 10.1523/JNEUROSCI.0820-22.2022 article palmisano_gait_2022 Bioengineering 2022 9 11 639 Freezing of gait (FOG) is a sudden episodic inability to produce effective stepping despite the intention to walk. It typically occurs during gait initiation (GI) or modulation and may lead to falls. We studied the anticipatory postural adjustments (imbalance, unloading, and stepping phase) at GI in 23 patients with Parkinson’s disease (PD) and FOG (PDF), 20 patients with PD and no previous history of FOG (PDNF), and 23 healthy controls (HCs). Patients performed the task when off dopaminergic medications. The center of pressure (CoP) displacement and velocity during imbalance showed significant impairment in both PDNF and PDF, more prominent in the latter patients. Several measurements were specifically impaired in PDF patients, especially the CoP displacement along the anteroposterior axis during unloading. The pattern of segmental center of mass (SCoM) movements did not show differences between groups. The standing postural profile preceding GI did not correlate with outcome measurements. We have shown impaired motor programming at GI in Parkinsonian patients. The more prominent deterioration of unloading in PDF patients might suggest impaired processing and integration of somatosensory information subserving GI. The unaltered temporal movement sequencing of SCoM might indicate some compensatory cerebellar mechanisms triggering time-locked models of body mechanics in PD. 8.4, 8.44 https://www.mdpi.com/2306-5354/9/11/639 2023-08-22 en 2306-5354 10.3390/bioengineering9110639 ChiaraPalmisano LauraBeccaria StefanHaufe JensVolkmann GianniPezzoli Ioannis U.Isaias article oala_machine_2021 Journal of Medical Systems 2021 45 12 105 Abstract Developers proposing new machine learning for health (ML4H) tools often pledge to match or even surpass the performance of existing tools, yet the reality is usually more complicated. Reliable deployment of ML4H to the real world is challenging as examples from diabetic retinopathy or Covid-19 screening show. We envision an integrated framework of algorithm auditing and quality control that provides a path towards the effective and reliable application of ML systems in healthcare. In this editorial, we give a summary of ongoing work towards that vision and announce a call for participation to the special issue  Machine Learning for Health: Algorithm Auditing &amp; Quality Control in this journal to advance the practice of ML4H auditing. 8.4, 8.44 https://link.springer.com/10.1007/s10916-021-01783-y 2022-09-13 en 0148-5598, 1573-689X 10.1007/s10916-021-01783-y LuisOala Andrew G.Murchison PradeepBalachandran ShrutiChoudhary JanaFehr Alixandro WerneckLeite Peter G.Goldschmidt ChristianJohner Elora D. M.Schörverth RoseNakasi MartinMeyer FedericoCabitza PatBaird CarolinPrabhu EvaWeicken XiaoxuanLiu MarkusWenzel SteffenVogler DarlingtonAkogo ShadaAlsalamah EmreKazim AdrianoKoshiyama SvenPiechottka SheenaMacpherson IanShadforth ReginaGeierhofer ChristianMatek JoachimKrois BrunoSanguinetti MatthewArentz PavolBielik SaulCalderon-Ramirez AussAbbood NicolasLanger StefanHaufe FerathKherif SameerPujari WojciechSamek ThomasWiegand article hashemi_unification_2021 NeuroImage 2021 239 118309 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S1053811921005851 2022-09-13 en 10538119 10.1016/j.neuroimage.2021.118309 AliHashemi ChangCai GittaKutyniok Klaus-RobertMüller Srikantan S.Nagarajan StefanHaufe article cai_robust_2021 NeuroImage 2021 225 117411 8.4, 8.44 https://linkinghub.elsevier.com/retrieve/pii/S105381192030896X 2022-09-13 en 10538119 10.1016/j.neuroimage.2020.117411 ChangCai AliHashemi MithunDiwakar StefanHaufe KensukeSekihara Srikantan S.Nagarajan article lichtner_predicting_2021 Scientific Reports 2021 11 1 13205 Abstract In a pandemic with a novel disease, disease-specific prognosis models are available only with a delay. To bridge the critical early phase, models built for similar diseases might be applied. To test the accuracy of such a knowledge transfer, we investigated how precise lethal courses in critically ill COVID-19 patients can be predicted by a model trained on critically ill non-COVID-19 viral pneumonia patients. We trained gradient boosted decision tree models on 718 (245 deceased) non-COVID-19 viral pneumonia patients to predict individual ICU mortality and applied it to 1054 (369 deceased) COVID-19 patients. Our model showed a significantly better predictive performance (AUROC 0.86 [95% CI 0.86–0.87]) than the clinical scores APACHE2 (0.63 [95% CI 0.61–0.65]), SAPS2 (0.72 [95% CI 0.71–0.74]) and SOFA (0.76 [95% CI 0.75–0.77]), the COVID-19-specific mortality prediction models of Zhou (0.76 [95% CI 0.73–0.78]) and Wang (laboratory: 0.62 [95% CI 0.59–0.65]; clinical: 0.56 [95% CI 0.55–0.58]) and the 4C COVID-19 Mortality score (0.71 [95% CI 0.70–0.72]). We conclude that lethal courses in critically ill COVID-19 patients can be predicted by a machine learning model trained on non-COVID-19 patients. Our results suggest that in a pandemic with a novel disease, prognosis models built for similar diseases can be applied, even when the diseases differ in time courses and in rates of critical and lethal courses. 8.4, 8.44 http://www.nature.com/articles/s41598-021-92475-7 2022-09-13 en 2045-2322 10.1038/s41598-021-92475-7 GregorLichtner FelixBalzer StefanHaufe NiklasGiesa FridtjofSchiefenhövel MalteSchmieding CarloJurth WolfgangKopp AltunaAkalin Stefan J.Schaller SteffenWeber-Carstens ClaudiaSpies Falkvon Dincklage inproceedings ali_hashemi_efficient_2021 2021 34 24855--24870 8.4, 8.44 https://proceedings.neurips.cc/paper/2021/file/d03a857a23b5285736c4d55e0bb067c8-Paper.pdf <prt>M. Ranzato</prt> and <prt>A. Beygelzimer</prt> and <prt>Y. Dauphin</prt> and <prt>P.S. Liang</prt> and <prt>J. Wortman Vaughan</prt> Curran Associates, Inc. Advances in <prt>Neural</prt> <prt>Information</prt> <prt>Processing</prt> <prt>Systems</prt> <prt>AliHashemi</prt> YijingGao ChangCai SanjayGhosh Klaus-RobertM<prt>\</prt>"<prt>u</prt>ller SrikantanNagarajan StefanHaufe