Constructing transferable and interpretable machine learning models for black carbon concentrations

Black carbon (BC) has received increasing attention from researchers due to its adverse health effects. However, in-situ BC measurements are often not included as a regulated variable in air quality monitoring networks. Machine learning (ML) models have been studied extensively to serve as virtual sensors to complement the reference instruments. This study evaluates and compares three white-box (WB) and four black-box (BB) ML models to estimate BC concentrations, with the focus to show their transferability and interpretability. We train the models with the long-term air pollutant and weather measurements in Barcelona urban background site, and test them in other European urban and traffic sites. Despite the difference in geographical locations and measurement sites, BC correlates the strongest with particle number concentration of accumulation mode (PNacc, r = 0.73-0.85) and nitrogen dioxide (NO2, r = 0.68-0.85) and the weakest with meteorological parameters. Due to its similarity of correlation behaviour, the ML models trained in Barcelona performs prominently at the traffic site in Helsinki (R2 = 0.80-0.86; mean absolute error MAE = 3.90-4.73 %) and at the urban background site in Dresden (R2 = 0.79-0.84; MAE = 4.23-4.82 %). WB models appear to explain less variability of BC than BB models, long short-term memory (LSTM) model of which outperforms the rest of the models. In terms of interpretability, we adopt several methods for individual model to quantify and normalize the relative importance of each input feature. The overall static relative importance commonly used for WB models demonstrate varying results from the dynamic values utilized to show local contribution used for BB models. PNacc and NO2 on average have the strongest absolute static contribution; however, they simultaneously impact the estimation positively and negatively at different sites. This comprehensive analysis demonstrates that the possibility of these interpretable air pollutant ML models to be transfered across space and time.

Virtual MEG sensors based on beamformer and independent component analysis can reconstruct epileptic activity as measured on simultaneous intracerebral recordings

The prevailing gold standard for presurgical determination of epileptogenic brain networks is intracerebral EEG, a potent yet invasive approach. Magnetoencephalography (MEG) is a state-of-the art non-invasive method for investigating epileptiform discharges. However, it is not clear at what level the precision offered by MEG can reach that of SEEG. Here, we present a strategy for non-invasively retrieving the constituents of the interictal network, with high spatial and temporal precision. Our method is based on MEG and a combination of spatial filtering and independent component analysis (ICA). We validated this approach in twelve patients with drug-resistant focal epilepsy, thanks to the unprecedented ground truth provided by simultaneous recordings of MEG and SEEG. A minimum variance adaptive beamformer estimated the source time series and ICA was used to further decompose these time series into network constituents (MEG-ICs), each having a time series (virtual electrode) and a topography (spatial distribution of amplitudes in the brain). We show that MEG has a considerable sensitivity of 0.80 and 0.84 and a specificity of 0.93 and 0.91 for reconstructing deep and superficial sources, respectively, when compared to the ground truth (SEEG). For each epileptic MEG-IC (n = 131), we found at least one significantly correlating SEEG contact close to zero lag after correcting for multiple comparisons. All the patients except one had at least one epileptic component that was highly correlated (Spearman rho>0.3) with that of SEEG traces. MEG-ICs correlated well with SEEG traces. The strength of correlation coefficients did not depend on the depth of the SEEG contacts or the clinical outcome of the patient. A significant proportion of the MEG-ICs (n = 83/131) were localized in proximity with their maximally correlating SEEG, within a mean distance of 20±12.18mm. Our research is the first to validate the MEG-retrieved beamformer IC sources against SEEG-derived ground truth in a simultaneous MEG-SEEG framework. Observations from the present study suggest that non-invasive MEG source components may potentially provide additional information, comparable to SEEG in a number of instances.

Virtual implantation using conventional scalp EEG delineates seizure onset and predicts surgical outcome in children with epilepsy

OBJECTIVE

Delineation of the seizure onset zone (SOZ) is required in children with drug resistant epilepsy (DRE) undergoing neurosurgery. Intracranial EEG (icEEG) serves as gold standard but has limitations. Here, we examine the utility of virtual implantation with electrical source imaging (ESI) on ictal scalp EEG for mapping the SOZ and predict surgical outcome.

METHODS

We retrospectively analyzed EEG data from 35 children with DRE who underwent surgery and dichotomized into seizure-free (SF) and non-seizure-free (NSF). We estimated virtual sensors (VSs) at brain locations that matched icEEG implantation and compared ictal patterns at VSs vs icEEG. We calculated the agreement between VSs SOZ and clinically defined SOZ and built receiver operating characteristic (ROC) curves to test whether it predicted outcome.

RESULTS

Twenty-one patients were SF after surgery. Moderate agreement between virtual and icEEG patterns was observed (kappa = 0.45, p < 0.001). Virtual SOZ agreement with clinically defined SOZ was higher in SF vs NSF patients (66.6% vs 41.6%, p = 0.01). Anatomical concordance of virtual SOZ with clinically defined SOZ predicted outcome (AUC = 0.73; 95% CI: 0.57-0.89; sensitivity = 66.7%; specificity = 78.6%; accuracy = 71.4%).

CONCLUSIONS

Virtual implantation on ictal scalp EEG can approximate the SOZ and predict outcome.

SIGNIFICANCE

SOZ mapping with VSs may contribute to tailoring icEEG implantation and predict outcome.

Fault diagnosis of an automated guided vehicle with torque and motion forces estimation: A case study

The paper is devoted to developing a new fault detection scheme for an Automated Guided Vehicle (AGV) on the basis of so-called virtual sensors (VSs) which provide the information regarding the current status of a vehicle. This set contains the estimates of lateral and longitudinal forces as well as the torque. The paper proposes a novel robust VSs design scheme which yields such estimates taking into account inevitable disturbances/noise and modelling uncertainty without any knowledge about tire models used in the AGV. The obtained estimates are used to generate the residuals and to diagnose the current status of the vehicle. Finally, the paper shows the experimental results concerning the application of the developed methods to fault detection of the self-designed and constructed AGV.

Automatic detection and visualisation of MEG ripple oscillations in epilepsy

High frequency oscillations (HFOs, 80–500 Hz) in invasive EEG are a biomarker for the epileptic focus. Ripples (80–250 Hz) have also been identified in non-invasive MEG, yet detection is impeded by noise, their low occurrence rates, and the workload of visual analysis. We propose a method that identifies ripples in MEG through noise reduction, beamforming and automatic detection with minimal user effort. We analysed 15 min of presurgical resting-state interictal MEG data of 25 patients with epilepsy. The MEG signal-to-noise was improved by using a cross-validation signal space separation method, and by calculating ~ 2400 beamformer-based virtual sensors in the grey matter. Ripples in these sensors were automatically detected by an algorithm optimized for MEG. A small subset of the identified ripples was visually checked. Ripple locations were compared with MEG spike dipole locations and the resection area if available. Running the automatic detection algorithm resulted in on average 905 ripples per patient, of which on average 148 ripples were visually reviewed. Reviewing took approximately 5 min per patient, and identified ripples in 16 out of 25 patients. In 14 patients the ripple locations showed good or moderate concordance with the MEG spikes. For six out of eight patients who had surgery, the ripple locations showed concordance with the resection area: 4/5 with good outcome and 2/3 with poor outcome. Automatic ripple detection in beamformer-based virtual sensors is a feasible non-invasive tool for the identification of ripples in MEG. Our method requires minimal user effort and is easily applicable in a clinical setting.

•

Cross-validation signal space separation and beamformer increase the SNR in MEG.

•

Automatic detection of MEG ripples in the time domain is feasible.

•

Our method identifies ripples with minimal user effort and is clinically applicable.

•

Automatically detected ripples are concordant with MEG spikes in 14/16 patients.

•

Automatically detected ripples are concordant with resection area in 6/8 patients.