Unraveling the physiological and psychosocial signatures of pain by machine learning.

BACKGROUND: Pain is a complex subjective experience, strongly impacting health and quality of life. Despite many attempts to find effective solutions, present treatments are generic, often unsuccessful, and present significant side effects. Designing individualized therapies requires understanding of multidimensional pain experience, considering physical and emotional aspects. Current clinical pain assessments, relying on subjective one-dimensional numeric self-reports, fail to capture this complexity. METHODS: To this aim, we exploited machine learning to disentangle physiological and psychosocial components shaping the pain experience. Clinical, psychosocial, and physiological data were collected from 118 chronic pain and healthy participants undergoing 40 pain trials (4,697 trials). FINDINGS: To understand the objective response to nociception, we classified pain from the physiological signals (accuracy >0.87), extracting the most important biomarkers. Then, using multilevel mixed-effects models, we predicted the reported pain, quantifying the mismatch between subjective level and measured physiological response. From these models, we introduced two metrics: TIP (subjective index of pain) and Φ (physiological index). These represent possible added value in the clinical process, capturing psychosocial and physiological pain dimensions, respectively. Patients with high TIP are characterized by frequent sick leave from work and increased clinical depression and anxiety, factors associated with long-term disability and poor recovery, and are indicated for alternative treatments, such as psychological ones. By contrast, patients with high Φ show strong nociceptive pain components and could benefit more from pharmacotherapy. CONCLUSIONS: TIP and Φ, explaining the multidimensionality of pain, might provide a new tool potentially leading to targeted treatments, thereby reducing the costs of inefficient generic therapies. FUNDING: RESC-PainSense, SNSF-MOVE-IT197271.

Towards enhanced functionality of vagus neuroprostheses through in silico optimized stimulation.

Bioelectronic therapies modulating the vagus nerve are promising for cardiovascular, inflammatory, and mental disorders. Clinical applications are however limited by side-effects such as breathing obstruction and headache caused by non-specific stimulation. To design selective and functional stimulation, we engineered VaStim, a realistic and efficient in-silico model. We developed a protocol to personalize VaStim in-vivo using simple muscle responses, successfully reproducing experimental observations, by combining models with trials conducted on five pigs. Through optimized algorithms, VaStim simulated the complete fiber population in minutes, including often omitted unmyelinated fibers which constitute 80% of the nerve. The model suggested that all Aα-fibers across the nerve affect laryngeal muscle, while heart rate changes were caused by B-efferents in specific fascicles. It predicted that tripolar paradigms could reduce laryngeal activity by 70% compared to typically used protocols. VaStim may serve as a model for developing neuromodulation therapies by maximizing efficacy and specificity, reducing animal experimentation.

Automated calibration of somatosensory stimulation using reinforcement learning.

BACKGROUND: The identification of the electrical stimulation parameters for neuromodulation is a subject-specific and time-consuming procedure that presently mostly relies on the expertise of the user (e.g., clinician, experimenter, bioengineer). Since the parameters of stimulation change over time (due to displacement of electrodes, skin status, etc.), patients undergo recurrent, long calibration sessions, along with visits to the clinics, which are inefficient and expensive. To address this issue, we developed an automatized calibration system based on reinforcement learning (RL) allowing for accurate and efficient identification of the peripheral nerve stimulation parameters for somatosensory neuroprostheses. METHODS: We developed an RL algorithm to automatically select neurostimulation parameters for restoring sensory feedback with transcutaneous electrical nerve stimulation (TENS). First, the algorithm was trained offline on a dataset comprising 49 subjects. Then, the neurostimulation was then integrated with a graphical user interface (GUI) to create an intuitive AI-based mapping platform enabling the user to autonomously perform the sensation characterization procedure. We assessed the algorithm against the performance of both experienced and naïve and of a brute force algorithm (BFA), on 15 nerves from five subjects. Then, we validated the AI-based platform on six neuropathic nerves affected by distal sensory loss. RESULTS: Our automatized approach demonstrated the ability to find the optimal values of neurostimulation achieving reliable and comfortable elicited sensations. When compared to alternatives, RL outperformed the naïve and BFA, significantly decreasing the time for mapping and the number of delivered stimulation trains, while improving the overall quality. Furthermore, the RL algorithm showed performance comparable to trained experimenters. Finally, we exploited it successfully for eliciting sensory feedback in neuropathic patients. CONCLUSIONS: Our findings demonstrated that the AI-based platform based on a RL algorithm can automatically and efficiently calibrate parameters for somatosensory nerve stimulation. This holds promise to avoid experts' employment in similar scenarios, thanks to the merging between AI and neurotech. Our RL algorithm has the potential to be used in other neuromodulation fields requiring a mapping process of the stimulation parameters. TRIAL REGISTRATION: ClinicalTrial.gov (Identifier: NCT04217005).

Multiparametric non-linear TENS modulation to integrate intuitive sensory feedback.

Objective. Transcutaneous electrical nerve stimulation (TENS) has been recently introduced in neurorehabilitation and neuroprosthetics as a promising, non-invasive sensory feedback restoration alternative to implantable neurostimulation. Yet, the adopted stimulation paradigms are typically based on single-parameter modulations (e.g. pulse amplitude (PA), pulse-width (PW) or pulse frequency (PF)). They elicit artificial sensations characterized by a low intensity resolution (e.g. few perceived levels), low naturalness and intuitiveness, hindering the acceptance of this technology. To address these issues, we designed novel multiparametric stimulation paradigms, featuring the simultaneous modulation of multiple parameters, and implemented them in real-time tests of performance when exploited as artificial sensory inputs.Approach. We initially investigated the contribution of PW and PF variations to the perceived sensation magnitude through discrimination tests. Then, we designed three multiparametric stimulation paradigms comparing them with a standard PW linear modulation in terms of evoked sensation naturalness and intensity. The most performant paradigms were then implemented in real-time in a Virtual Reality-TENS platform to assess their ability to provide intuitive somatosensory feedback in a functional task.Main results. Our study highlighted a strong negative correlation between perceived naturalness and intensity: less intense sensations are usually deemed as more similar to natural touch. In addition, we observed that PF and PW changes have a different weight on the perceived sensation intensity. As a result, we adapted the activation charge rate (ACR) equation, proposed for implantable neurostimulation to predict the perceived intensity while co-modulating the PF and charge per pulse, to TENS (ACRT). ACRTallowed to design different multiparametric TENS paradigms with the same absolute perceived intensity. Although not reported as more natural, the multiparametric paradigm, based on sinusoidal PF modulation, resulted being more intuitive and subconsciously integrated than the standard linear one. This allowed subjects to achieve a faster and more accurate functional performance.Significance. Our findings suggest that TENS-based, multiparametric neurostimulation, despite not consciously perceived naturally, can provide integrated and more intuitive somatosensory information, as functionally proved. This could be exploited to design novel encoding strategies able to improve the performance of non-invasive sensory feedback technologies.

Explainable domain transfer of distant supervised cancer subtyping model via imaging-based rules extraction.

Image texture analysis has for decades represented a promising opportunity for cancer assessment and disease progression evaluation, evolving in a discipline, i.e., radiomics. However, the road to a complete translation into clinical practice is still hampered by intrinsic limitations. As purely supervised classification models fail in devising robust imaging-based biomarkers for prognosis, cancer subtyping approaches would benefit from the employment of distant supervision, for instance exploiting survival/recurrence information. In this work, we assessed, tested, and validated the domain-generality of our previously proposed Distant Supervised Cancer Subtyping model on Hodgkin Lymphoma. We evaluate the model performance on two independent datasets coming from two hospitals, comparing and analyzing the results. Although successful and consistent, the comparison confirmed the instability of radiomics due to an across-center lack of reproducibility, leading to explainable results in one center and poor interpretability in the other. We thus propose a Random Forest-based Explainable Transfer Model for testing the domain-invariance of imaging biomarkers extracted from retrospective cancer subtyping. In doing so, we tested the predictive ability of cancer subtyping in a validation and perspective setting, which led to successful results and supported the domain-generality of the proposed approach. On the other hand, the extraction of decision rules enables to draw of risk factors and robust biomarkers to inform clinical decisions. This work shows the potentialities of the Distant Supervised Cancer Subtyping model to be further evaluated in larger multi-center datasets, to reliably translate radiomics into medical practice. The code is available at this GitHub repository.

The Development of an Intelligent Agent to Detect and Non-Invasively Characterize Lung Lesions on CT Scans: Ready for the "Real World"?

(1) Background: Once lung lesions are identified on CT scans, they must be characterized by assessing the risk of malignancy. Despite the promising performance of computer-aided systems, some limitations related to the study design and technical issues undermine these tools' efficiency; an "intelligent agent" to detect and non-invasively characterize lung lesions on CT scans is proposed. (2) Methods: Two main modules tackled the detection of lung nodules on CT scans and the diagnosis of each nodule into benign and malignant categories. Computer-aided detection (CADe) and computer aided-diagnosis (CADx) modules relied on deep learning techniques such as Retina U-Net and the convolutional neural network; (3) Results: Tests were conducted on one publicly available dataset and two local datasets featuring CT scans acquired with different devices to reveal deep learning performances in "real-world" clinical scenarios. The CADe module reached an accuracy rate of 78%, while the CADx's accuracy, specificity, and sensitivity stand at 80%, 73%, and 85.7%, respectively; (4) Conclusions: Two different deep learning techniques have been adapted for CADe and CADx purposes in both publicly available and private CT scan datasets. Experiments have shown adequate performance in both detection and diagnosis tasks. Nevertheless, some drawbacks still characterize the supervised learning paradigm employed in networks such as CNN and Retina U-Net in real-world clinical scenarios, with CT scans from different devices with different sensors' fingerprints and spatial resolution. Continuous reassessment of CADe and CADx's performance is needed during their implementation in clinical practice.

Image Embeddings Extracted from CNNs Outperform Other Transfer Learning Approaches in Classification of Chest Radiographs.

To identify the best transfer learning approach for the identification of the most frequent abnormalities on chest radiographs (CXRs), we used embeddings extracted from pretrained convolutional neural networks (CNNs). An explainable AI (XAI) model was applied to interpret black-box model predictions and assess its performance. Seven CNNs were trained on CheXpert. Three transfer learning approaches were thereafter applied to a local dataset. The classification results were ensembled using simple and entropy-weighted averaging. We applied Grad-CAM (an XAI model) to produce a saliency map. Grad-CAM maps were compared to manually extracted regions of interest, and the training time was recorded. The best transfer learning model was that which used image embeddings and random forest with simple averaging, with an average AUC of 0.856. Grad-CAM maps showed that the models focused on specific features of each CXR. CNNs pretrained on a large public dataset of medical images can be exploited as feature extractors for tasks of interest. The extracted image embeddings contain relevant information that can be used to train an additional classifier with satisfactory performance on an independent dataset, demonstrating it to be the optimal transfer learning strategy and overcoming the need for large private datasets, extensive computational resources, and long training times.

Preoperative [11C]methionine PET to personalize treatment decisions in patients with lower-grade gliomas.

BACKGROUND: PET with radiolabeled amino acids is used in the preoperative evaluation of patients with glial neoplasms. This study aimed to assess the role of [11C]methionine (MET) PET in assessing molecular features, tumor extent, and prognosis in newly diagnosed lower-grade gliomas (LGGs) surgically treated. METHODS: One hundred and fifty-three patients with a new diagnosis of grade 2/3 glioma who underwent surgery at our Institution and were imaged preoperatively using [11C]MET PET/CT were retrospectively included. [11C]MET PET images were qualitatively and semi-quantitatively analyzed using tumor-to-background ratio (TBR). Progression-free survival (PFS) rates were estimated using the Kaplan-Meier method and Cox proportional-hazards regression was used to test the association of clinicopathological and imaging data to PFS. RESULTS: Overall, 111 lesions (73%) were positive, while thirty-two (21%) and ten (6%) were isometabolic and hypometabolic at [11C]MET PET, respectively. [11C]MET uptake was more common in oligodendrogliomas than IDH-mutant astrocytomas (87% vs 50% of cases, respectively). Among [11C]MET-positive gliomas, grade 3 oligodendrogliomas had the highest median TBRmax (3.22). In 25% of patients, PET helped to better delineate tumor margins compared to MRI only. In IDH-mutant astrocytomas, higher TBRmax values at [11C]MET PET were independent predictors of shorter PFS. CONCLUSIONS: This work highlights the role of preoperative [11C]MET PET in estimating the type of suspected LGGs, assessing tumor extent, and predicting biological behavior and prognosis of histologically confirmed LGGs. Our findings support the implementation of [11C]MET PET in routine clinical practice to better manage these neoplasms.

State-of-the-art of FAPI-PET imaging: a systematic review and meta-analysis.

INTRODUCTION: Fibroblast activation protein-α (FAPα) is overexpressed on cancer-associated fibroblasts in approximately 90% of epithelial neoplasms, representing an appealing target for therapeutic and molecular imaging applications. [68 Ga]Ga-labelled radiopharmaceuticals-FAP-inhibitors (FAPI)-have been developed for PET. We systematically reviewed and meta-analysed published literature to provide an overview of its clinical role. MATERIALS AND METHODS: The search, limited to January 1st, 2018-March 31st, 2021, was performed on MedLine and Embase databases using all the possible combinations of terms "FAP", "FAPI", "PET/CT", "positron emission tomography", "fibroblast", "cancer-associated fibroblasts", "CAF", "molecular imaging", and "fibroblast imaging". Study quality was assessed using the QUADAS-2 criteria. Patient-based and lesion-based pooled sensitivities/specificities of FAPI PET were computed using a random-effects model directly from the STATA "metaprop" command. Between-study statistical heterogeneity was tested (I2-statistics). RESULTS: Twenty-three studies were selected for systematic review. Investigations on staging or restaging head and neck cancer (n = 2, 29 patients), abdominal malignancies (n = 6, 171 patients), various cancers (n = 2, 143 patients), and radiation treatment planning (n = 4, 56 patients) were included in the meta-analysis. On patient-based analysis, pooled sensitivity was 0.99 (95% CI 0.97-1.00) with negligible heterogeneity; pooled specificity was 0.87 (95% CI 0.62-1.00), with negligible heterogeneity. On lesion-based analysis, sensitivity and specificity had high heterogeneity (I2 = 88.56% and I2 = 97.20%, respectively). Pooled sensitivity for the primary tumour was 1.00 (95% CI 0.98-1.00) with negligible heterogeneity. Pooled sensitivity/specificity of nodal metastases had high heterogeneity (I2 = 89.18% and I2 = 95.74%, respectively). Pooled sensitivity in distant metastases was good (0.93 with 95% CI 0.88-0.97) with negligible heterogeneity. CONCLUSIONS: FAPI-PET appears promising, especially in imaging cancers unsuitable for [18F]FDG imaging, particularly primary lesions and distant metastases. However, high-level evidence is needed to define its role, specifically to identify cancer types, non-oncological diseases, and clinical settings for its applications.

Radiomics and gene expression profile to characterise the disease and predict outcome in patients with lung cancer.

OBJECTIVE: The objectives of our study were to assess the association of radiomic and genomic data with histology and patient outcome in non-small cell lung cancer (NSCLC). METHODS: In this retrospective single-centre observational study, we selected 151 surgically treated patients with adenocarcinoma or squamous cell carcinoma who performed baseline [18F] FDG PET/CT. A subgroup of patients with cancer tissue samples at the Institutional Biobank (n = 74/151) was included in the genomic analysis. Features were extracted from both PET and CT images using an in-house tool. The genomic analysis included detection of genetic variants, fusion transcripts, and gene expression. Generalised linear model (GLM) and machine learning (ML) algorithms were used to predict histology and tumour recurrence. RESULTS: Standardised uptake value (SUV) and kurtosis (among the PET and CT radiomic features, respectively), and the expression of TP63, EPHA10, FBN2, and IL1RAP were associated with the histotype. No correlation was found between radiomic features/genomic data and relapse using GLM. The ML approach identified several radiomic/genomic rules to predict the histotype successfully. The ML approach showed a modest ability of PET radiomic features to predict relapse, while it identified a robust gene expression signature able to predict patient relapse correctly. The best-performing ML radiogenomic rule predicting the outcome resulted in an area under the curve (AUC) of 0.87. CONCLUSIONS: Radiogenomic data may provide clinically relevant information in NSCLC patients regarding the histotype, aggressiveness, and progression. Gene expression analysis showed potential new biomarkers and targets valuable for patient management and treatment. The application of ML allows to increase the efficacy of radiogenomic analysis and provides novel insights into cancer biology.

Distributed learning: a reliable privacy-preserving strategy to change multicenter collaborations using AI.

PURPOSE: The present scoping review aims to assess the non-inferiority of distributed learning over centrally and locally trained machine learning (ML) models in medical applications. METHODS: We performed a literature search using the term "distributed learning" OR "federated learning" in the PubMed/MEDLINE and EMBASE databases. No start date limit was used, and the search was extended until July 21, 2020. We excluded articles outside the field of interest; guidelines or expert opinion, review articles and meta-analyses, editorials, letters or commentaries, and conference abstracts; articles not in the English language; and studies not using medical data. Selected studies were classified and analysed according to their aim(s). RESULTS: We included 26 papers aimed at predicting one or more outcomes: namely risk, diagnosis, prognosis, and treatment side effect/adverse drug reaction. Distributed learning was compared to centralized or localized training in 21/26 and 14/26 selected papers, respectively. Regardless of the aim, the type of input, the method, and the classifier, distributed learning performed close to centralized training, but two experiments focused on diagnosis. In all but 2 cases, distributed learning outperformed locally trained models. CONCLUSION: Distributed learning resulted in a reliable strategy for model development; indeed, it performed equally to models trained on centralized datasets. Sensitive data can get preserved since they are not shared for model development. Distributed learning constitutes a promising solution for ML-based research and practice since large, diverse datasets are crucial for success.