Real-time endoscopic rectal lesion characterization employing artificial intelligence (AI) and near-infrared (NIR) imaging of the fluorescence perfusion indicator agent Indocyanine Green (ICG) has demonstrated promise. However, commercially available fluorescence endoscopes do not possess the flexibility and anatomical reach capabilities of colonoscopy while commercial flexible scopes do not yet provide beyond visible spectral imaging. This limits the application of this AI-NIR classification technology. Here, to close this technical gap, we present our development of a colonoscope-compatible flexible imaging probe for NIR-ICG visualization combined with a full field of view machine learning (ML) algorithm for fluorescence quantification and perfusion pattern cross-correlation (including first in human testing). The imaging probe is capable of 133µm minimum object resolution, with a maximum working distance of 50mm and an excitation illumination power of 52mW with 75o average field of illumination (meaning minimum device tip distance from target is 13 mm for a 2 cm polyp). The system demonstrated ex-vivo and in-vivo NIR visualization of clinically relevant concentrations of ICG in both resected and in situ (extracorporeally) colon in patients undergoing colorectal resection. A previously developed AI-NIR perfusion quantification algorithm was applied to videos of a bench model of varying ICG flow captured with the developed flexible system with added ML features generated full field of view pixel-level fluorescence time-series measurements capable of distinguishing distinct ICG flow regions in the image via correlative dynamic fluorescence intensity profiles. Jaccard Index comparison of the AI -generated flow regions against manually delineated flow regions resulted in 79% accuracy. While further clinical validation of the AI-NIR polyp classification method is on-going (in the Horizon Europe Awarded CLASSICA project), other use case applications of NIR colonoscopy include simpler perioperative perfusion assessment in patients undergoing colorectal resection and combination with targeted agents in development thus encouraging continuing development and design optimization of this flexible NIR imaging probe to enable clinical and commercial translation.

Computational competitions are the standard for benchmarking medical image analysis algorithms, but they typically use small curated test datasets acquired at a few centers, leaving a gap to the reality of diverse multicentric patient data. To this end, the Federated Tumor Segmentation (FeTS) Challenge represents the paradigm for real-world algorithmic performance evaluation. The FeTS challenge is a competition to benchmark (i) federated learning aggregation algorithms and (ii) state-of-the-art segmentation algorithms, across multiple international sites. Weight aggregation and client selection techniques were compared using a multicentric brain tumor dataset in realistic federated learning simulations, yielding benefits for adaptive weight aggregation, and efficiency gains through client sampling. Quantitative performance evaluation of state-of-the-art segmentation algorithms on data distributed internationally across 32 institutions yielded good generalization on average, albeit the worst-case performance revealed data-specific modes of failure. Similar multi-site setups can help validate the real-world utility of healthcare AI algorithms in the future.

Purpose Perioperative decision making for large (> 2 cm) rectal polyps with ambiguous features is complex. The most common intraprocedural assessment is clinician judgement alone while radiological and endoscopic biopsy can provide periprocedural detail. Fluorescence-augmented machine learning (FA-ML) methods may optimise local treatment strategy. Methods Surgeons of varying grades, all performing colonoscopies independently, were asked to visually judge endoscopic videos of large benign and early-stage malignant (potentially suitable for local excision) rectal lesions on an interactive video platform (Mindstamp) with results compared with and between final pathology, radiology and a novel FA-ML classifier. Statistical analyses of data used Fleiss Multi-rater Kappa scoring, Spearman Coefficient and Frequency tables. Results Thirty-two surgeons judged 14 ambiguous polyp videos (7 benign, 7 malignant). In all cancers, initial endoscopic biopsy had yielded false-negative results. Five of each lesion type had had a pre-excision MRI with a 60% false-positive malignancy prediction in benign lesions and a 60% over-staging and 40% equivocal rate in cancers. Average clinical visual cancer judgement accuracy was 49% (with only 'fair' inter-rater agreement), many reporting uncertainty and higher reported decision confidence did not correspond to higher accuracy. This compared to 86% ML accuracy. Size was misjudged visually by a mean of 20% with polyp size underestimated in 4/6 and overestimated in 2/6. Subjective narratives regarding decision-making requested for 7/14 lesions revealed wide rationale variation between participants. Conclusion Current available clinical means of ambiguous rectal lesion assessment is suboptimal with wide inter-observer variation. Fluorescence based AI augmentation may advance this field via objective, explainable ML methods.

Fluorescence-guided oncology promises to improve both the detection and treatment of malignancy. We sought to investigate the temporal distribution of indocyanine green (ICG), an exogenous fluorophore in human colorectal cancer. This analysis aims to enhance our understanding of ICG’s effectiveness in current tumour detection and inform potential future diagnostic and therapeutic enhancements. Methods: Fifty consenting patients undergoing treatment for suspected/confirmed colorectal neoplasia provided near infrared (NIR) video and imagery of transanally recorded and ex vivo resected rectal lesions following intravenous ICG administration (0.25 mg/kg), with a subgroup providing tissue samples for microscopic (including near infrared) analysis. Computer vision techniques detailed macroscopic ‘early’ (<15 min post ICG administration) and ‘late’ (>2 h) tissue fluorescence appearances from surgical imagery with digital NIR scanning (Licor, Lincoln, NE, USA) and from microscopic analysis (Nikon, Tokyo, Japan) undertaken by a consultant pathologist detailing tissue-level fluorescence distribution over the same time. Results: Significant intra-tumoural fluorescence heterogeneity was seen ‘early’ in malignant versus benign lesions. In all ‘early’ samples, fluorescence was predominantly within the tissue stroma, with uptake within plasma cells, blood vessels and lymphatics, but not within malignant or healthy glands. At ‘late’ stage observation, fluorescence was visualised non-uniformly within the intracellular cytoplasm of malignant tissue but not retained in benign glands. Fluorescence also accumulated within any present peritumoural inflammatory tissue. Conclusion: This study demonstrates the time course diffusion patterns of ICG through both benign and malignant tumours in vivo in human patients at both macroscopic and microscopic levels, demonstrating important cellular drivers and features of geolocalisation and how they differ longitudinally after exposure to ICG.

Intraoperative indocyanine green fluorescence angiography (ICGFA) aims to reduce colorectal anastomotic complications. However, signal interpretation is inconsistent and confounded by patient physiology and system behaviours. Here, we demonstrate a proof of concept of a novel clinical and computational method for patient calibrated quantitative ICGFA (QICGFA) bowel transection recommendation. Patients undergoing elective colorectal resection had colonic ICGFA both immediately after operative commencement prior to any dissection and again, as usual, just before anastomotic construction. Video recordings of both ICGFA acquisitions were blindly quantified post hoc across selected colonic regions of interest (ROIs) using tracking-quantification software and computationally compared with satisfactory perfusion assumed in second time-point ROIs, demonstrating 85

Introduction: Intraoperative indocyanine green fluorescence angiography (ICGFA) perfusion assessment has been demonstrated to reduce complications in reconstructive surgery. This study sought to advance ICGFA flap perfusion assessment via quantification methodologies. Method: Patients undergoing pedicled and free flap reconstruction were subjected to intraoperative ICGFA flap perfusion assessment using either an open or endoscopic system. Patient demographics, clinical impact of ICGFA and outcomes were documented. From the ICGFA recordings, fluorescence signal quality, as well as inflow/outflow milestones for the flap and surrounding (control) tissue were computationally quantified post hoc and compared on a region of interest (ROI) level. Further software development intended full flap quantification, metric computation and heatmap generation. Results: Fifteen patients underwent ICGFA assessment at reconstruction (8 head and neck, 6 breast and 1 perineum) including 10 free and 5 pedicled flaps. Visual ICGFA interpretation altered on-table management in 33.3% of cases, with flap edges trimmed in 4 and a re-anastomosis in 1 patient. One patient suffered post-operative flap dehiscence. Laparoscopic camera use proved feasible but recorded a lower quality signal than the open system.Using established and novel metrics, objective ICGFA signal ROI quantification permitted perfusion comparisons between the flap and surrounding tissue. Full flap assessment feasibility was demonstrated by computing all pixels and subsequent outputs summarisation as heatmaps. Conclusion: This trial demonstrated the feasibility and potential for ICGFA with operator based and quantitative flap perfusion assessment across several reconstructive applications. Further development and implementation of these computational methods requires technique and device standardisation.

In this work, we devise robust and efficient learning protocols for orchestrating a Federated Learning (FL) process for the Federated Tumor Segmentation Challenge (FeTS 2022). Enabling FL for FeTS setup is challenging mainly due to data heterogeneity among collaborators and communication cost of training. To tackle these challenges, we propose a Robust Learning Protocol (RoLePRO) which is a combination of server-side adaptive optimisation (e.g., server-side Adam) and judicious parameter (weights) aggregation schemes (e.g., adaptive weighted aggregation). RoLePRO takes a two-phase approach, where the first phase consists of vanilla Federated Averaging, while the second phase consists of a judicious aggregation scheme that uses a sophisticated re-weighting, all in the presence of an adaptive optimisation algorithm at the server. We draw insights from extensive experimentation to tune learning rates for the two phases.

This paper introduces a novel methodology for Feature Selection for Functional Classification, FSFC, that addresses the challenge of jointly performing feature selection and classification of functional data in scenarios with categorical responses and multivariate longitudinal features. FSFC tackles a newly defined optimization problem that integrates logistic loss and functional features to identify the most crucial variables for classification. To address the minimization procedure, we employ functional principal components and develop a new adaptive version of the Dual Augmented Lagrangian algorithm. The computational efficiency of FSFC enables handling high-dimensional scenarios where the number of features may considerably exceed the number of statistical units. Simulation experiments demonstrate that FSFC outperforms other machine learning and deep learning methods in computational time and classification accuracy. Furthermore, the FSFC feature selection capability can be leveraged to significantly reduce the problem's dimensionality and enhance the performances of other classification algorithms. The efficacy of FSFC is also demonstrated through a real data application, analyzing relationships between four chronic diseases and other health and demographic factors.

Introduction: Fluorescence guided surgery for the identification of colorectal liver metastases (CRLM) can be bet -ter with low specificity and antecedent dosing impracticalities limiting indocyanine green (ICG) usefulness cur-rently. We investigated the application of artificial intelligence methods (AIM) to demonstrate and characterise CLRMs based on dynamic signalling immediately following intraoperative ICG administration.Methods: Twenty-five patients with liver surface lesions (24 CRLM and 1 benign cyst) undergoing open/laparo-scopic/robotic procedures were studied. ICG (0.05 mg/kg) was administered with near-infrared recording of fluorescence perfusion. User-selected region-of-interest (ROI) perfusion profiles were generated, milestones re-lating to ICG inflow/outflow extracted and used to train a machine learning (ML) classifier. 2D heatmaps were constructed in a subset using AIM to depict whole screen imaging based on dynamic tissue-ICG interaction. Fluo-rescence appearances were also assessed microscopically (using H&E and fresh-frozen preparations) to provide tissue-level explainability of such methods.Results: The ML algorithm correctly classified 97.2 % of CRLM ROIs (n = 132) and all benign lesion ROIs (n = 6) within 90-s of ICG administration following initial mathematical curve analysis identifying ICG inflow/outflow differentials between healthy liver and CRLMs. Time-fluorescence plots extracted for each pixel in 10 lesions en-abled creation of 2D characterising heatmaps using flow parameters and through unsupervised ML. Microscopy confirmed statistically less CLRM fluorescence vs adjacent liver (mean +/- std deviation signal/area 2.46 +/- 9.56 vs 507.43 +/- 160.82 respectively p < 0.001) with H&E diminishing ICG signal (n = 4).Conclusion: ML accurately identifies CRLMs from surrounding liver tissue enabling representative 2D mapping of such lesions from their fluorescence perfusion patterns using AIM. This may assist in reducing positive margin rates at metastatectomy and in identifying unexpected/occult malignancies.(c) 2023 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http:// creativecommons.org/licenses/by/4.0/).

Introduction Indocyanine green (ICG) quantification and assessment by machine learning (ML) could discriminate tissue types through perfusion characterisation, including delineation of malignancy. Here, we detail the important challenges overcome before effective clinical validation of such capability in a prospective patient series of quantitative fluorescence angiograms regarding primary and secondary colorectal neoplasia. Methods ICG perfusion videos from 50 patients (37 with benign (13) and malignant (24) rectal tumours and 13 with colorectal liver metastases) of between 2- and 15-min duration following intravenously administered ICG were formally studied (clinicaltrials.gov: NCT04220242). Video quality with respect to interpretative ML reliability was studied observing practical, technical and technological aspects of fluorescence signal acquisition. Investigated parameters included ICG dosing and administration, distance–intensity fluorescent signal variation, tissue and camera movement (including real-time camera tracking) as well as sampling issues with user-selected digital tissue biopsy. Attenuating strategies for the identified problems were developed, applied and evaluated. ML methods to classify extracted data, including datasets with interrupted time-series lengths with inference simulated data were also evaluated. Results Definable, remediable challenges arose across both rectal and liver cohorts. Varying ICG dose by tissue type was identified as an important feature of real-time fluorescence quantification. Multi-region sampling within a lesion mitigated representation issues whilst distance–intensity relationships, as well as movement-instability issues, were demonstrated and ameliorated with post-processing techniques including normalisation and smoothing of extracted time–fluorescence curves. ML methods (automated feature extraction and classification) enabled ML algorithms glean excellent pathological categorisation results (AUC-ROC > 0.9, 37 rectal lesions) with imputation proving a robust method of compensation for interrupted time-series data with duration discrepancies. Conclusion Purposeful clinical and data-processing protocols enable powerful pathological characterisation with existing clinical systems. Video analysis as shown can inform iterative and definitive clinical validation studies on how to close the translation gap between research applications and real-world, real-time clinical utility.

Advances in microvascular surgical techniques have expanded the scope of free flap (FF) reconstruction. Tools predicting wound1 and overall2 outcomes now enhance operative planning, which can also be facilitated by 3D reconstructions. Perforator mapping and postanastomotic FF perfusion can be guided by Doppler ultrasound, thermographic, and hyperspectral imaging, with implantable probes also expediting postoperative FF compromise detection, allowing prompt operative/nonoperative flap salvage procedures,3 for example, venous supercharging. Despite this, complication rates in head and neck reconstruction remain higher (7.7%) than in other regions, for example, breast (5.11%) and extremities (1.32%).4 Furthermore, complications in this region are potentially catastrophic, exposing major vessels and potentially precipitating anastomotic leaks or salivary fistulae. Routine dual venous anastomosis (DVA)5 diminishes venous compromise and subsequent thrombosis, especially when anastomosing onto different venous systems, for example, the internal jugular and external jugular veins. However, DVA necessitates further dissection and longer operating time. Separately, indocyanine green fluorescence angiography (ICGFA) improves outcomes in autologous breast reconstruction. Here, we seek to demonstrate and quantify improved perfusion with DVA via metrics applied and validated in other surgical specialties. Within a trial (ClinicalTrials.gov Identifier: NCT04220242), operative ICGFA was performed on an anterolateral thigh (ALT) flap deployed to reconstruct an 8 × 10 cm defect following total laryngectomy and bilateral selective neck dissection for invasive squamous cell carcinoma (T3N0) in a 66-year-old man. The ALT flap was raised using a subfascial technique through the intermuscular septum and a standard pedicle, including a single artery and two veins, was harvested. The lateral circumflex femoral artery was anastomosed (end-to-end, 9-0 nylon) to the facial artery and both venae comitantes to the external jugular and internal jugular tributaries, respectively, using 2.5-mm couplers (Synovis). ICGFA (0.1 mg/kg ICG) was performed and visually interpreted following single-vein anastomosis and repeated following DVA using EleVision IR (Medtronic, Ireland). The flap healed, and the patient was discharged without any flap complications. Using bespoke software (IBM Research Europe, Ireland), a fluorescence-time series from every pixel was captured from the 5-minute-long videos recorded 30 minutes apart, following each anastomosis. From these curves, the maximum intensity (Fmax), rate of inflow (upslope), and rate of outflow (downslope) were determined, statistically interrogated, and computed into heat maps. Visual ICGFA assessment showed brisker and broader distribution following DVA versus single-vein anastomosis (Fig. 1). [See Video (online), which displays fluorescence angiography following 0.1 mg/kg of ICG using the EleVision IR (Medtronic, Ireland) near-infrared open camera system following single venous anastomosis (left) and DVA (right) on an anterolateral thigh microvascular FF.]Fig. 1.: Flap perfusion metrics calculated from the fluorescence video data (following single venous anastomosis and DVA) were illustrated as an augmented view over the camera displayed a white light image (top left) of the ALT flap. These metrics for every pixel of the flap were enumerated as a heat map with a color scale increasing upward from purple to yellow. Fmax (peak intensity, top right) in grayscale units (g.u.), inflow (upslope, bottom left), and outflow (downslope, more negative denotes faster outflow) gradients (g.u./s: second). The letters denote the target donor site of the flap tissue with P denoting the pharynx and S denoting the skin. {"href":"Single Video Player","role":"media-player-id","content-type":"play-in-place","position":"float","orientation":"portrait","label":"Video 1","caption":"displays fluorescence angiography following 0.1 mg/kg of ICG using the EleVision� IR, Medtronic, Ireland near-infrared open camera system following single (left) and dual (right) venous anastomosis on an anterolateral thigh microvascular free flap.","object-id":[{"pub-id-type":"doi","id":""},{"pub-id-type":"other","content-type":"media-stream-id","id":"1_olctn1kk"},{"pub-id-type":"other","content-type":"media-source","id":"Kaltura"}]} ICGFA metrics significantly favored improved perfusion following the DVA (Wilcoxon signed-rank P < 0.001, Table 1). The flap fluoresced brighter (Fmax 222.15 ± 15.57 versus 157.22 ± 51.42 g.u.: grayscale units) with steeper upslope (inflow 4.36 ± 1.98 versus 0.81 ± 0.72 g.u./s) and downslope (outflow −0.08 ± 0.03 versus −0.04 ± 0.03 g.u./s) gradients. Table 1. - Data Analysis for Full Flap Quantitative ICGFA following Single-vein Anastomosis and DVA: Including the Number of Pixels (n) at Which the Fluorescence Intensity Change over Time Was Converted into a Time Series, Mean ± SD for the Maximum Fluorescence (Fmax in g.u.: Grayscale Units) Inflow (Upslope) and Outflow (Downslope) Gradients (g.u./s: Second) Full Flap (per Pixel) Quantitative ICGFA Data following Single and Dual Venous Anastomosis Single Anastomosis, Mean ± SD Dual Anastomosis, Mean ± SD P Time fluorescence curves sampled per flap (n) 80,395 106,758 Fmax (g.u.) 157.22 ± 51.42 222.15 ± 15.57 <0.001* Upslope (inflow) gradient (g.u./s) 0.81 ± 0.72 4.36 ± 1.98 <0.001* Downslope (outflow) gradient (g.u./s) −0.04 ± 0.03 −0.08 ± 0.03 <0.001* P values denote statistical significance for Wilcoxon signed-rank test (following Shapiro-Wilk test for normality, SPSS version 27; IBM).*P < 0.05 These proposed methodologies might quantitatively support decision-making, removing interpretation variability. Furthermore, the selected camera compensates for interassessment distance-related fluorescence variations. Despite a brief interassessment period, flap flow is known to improve with time, even before choke vessel opening. Confounding variables require assessment using control cases in an investigative case series. This case report augments evidence supporting routine DVA in head and neck surgery when technically possible by objectively demonstrating improved perfusion via full flap computational ICGFA assessment. DISCLOSURES Dr. Cahill is named on a patent filed in relation to processes for visual determination of tissue biology, receives speaker fees from Stryker Corp, Ethicon/J&J, and Olympus, research funding from Intuitive Corp, consultancy fees from Arthrex, Diagnostic Green, Distalmotion, and Medtronic (Touch Surgery), and holds research funding from the Irish Government (DTIF) in collaboration with IBM Research Europe in Ireland, from EU Horizon 2020 in collaboration with Palliare and Steripak, from Horizon Europe in collaboration with Arctur, and from Intuitive and Medtonic for specific research and development awards. Jeffrey Dalli was employed as a researcher in the DTIF and is a recipient of the TESS scholarship (MALTA). Dr. Epperlein is a full-time employee of IBM Research Europe, a division of IBM, which provides technical products and services worldwide to government, healthcare, and life-sciences companies. The other authors have no financial interest to declare. ACKNOWLEDGMENT This research includes a consenting adult patient and was approved by the Mater Misericordiae University Hospital (Dublin, Ireland) institutional review board (Reference 1/378/2092, ClinicalTrials.gov Identifier: NCT04220242).

SignificanceAs clinical evidence on the colorectal application of indocyanine green (ICG) perfusion angiography accrues, there is also interest in computerizing decision support. However, user interpretation and software development may be impacted by system factors affecting the displayed near-infrared (NIR) signal.AimWe aim to assess the impact of camera positioning on the displayed NIR signal across different open and laparoscopic camera systems.ApproachThe effects of distance, movement, and target location (center versus periphery) on the displayed fluorescence signal of different systems were measured under electromagnetic stereotactic guidance from an ICG-albumin model and in vivo during surgery.ResultsSystems displayed distinct fluorescence performances with variance apparent with scope optical lens configuration (0 deg versus 30 deg), movement, target positioning, and distance. Laparoscopic system readings fitted inverse square function distance-intensity curves with one device and demonstrated a direction dependent sigmoid curve. Laparoscopic cameras presented central targets as brighter than peripheral ones, and laparoscopes with angled optical lens configurations had a diminished field of view. One handheld open system also showed a distance-intensity relationship, whereas the other maintained a consistent signal despite distance, but both presented peripheral targets brighter than central ones.ConclusionsOptimal clinical use and signal computational development requires detailed appreciation of system behaviors.

Delta like canonical notch ligand 4 (Dll4) expression levels in tumors are known to affect the efficacy of cancer therapies. This study aimed to develop a model to predict Dll4 expression levels in tumors using dynamic enhanced near-infrared (NIR) imaging with indocyanine green (ICG). Two rat-based consomic xenograft (CXM) strains of breast cancer with different Dll4 expression levels and eight congenic xenograft strains were studied. Principal component analysis (PCA) was used to visualize and segment tumors, and modified PCA techniques identified and analyzed tumor and normal regions of interest (ROIs). The average NIR intensity for each ROI was calculated from pixel brightness at each time interval, yielding easily interpretable features including the slope of initial ICG uptake, time to peak perfusion, and rate of ICG intensity change after reaching half-maximum intensity. Machine learning algorithms were applied to select discriminative features for classification, and model performance was evaluated with a confusion matrix, receiver operating characteristic curve, and area under the curve. The selected machine learning methods accurately identified host Dll4 expression alterations with sensitivity and specificity above 90%. This may enable stratification of patients for Dll4 targeted therapies. NIR imaging with ICG can noninvasively assess Dll4 expression levels in tumors and aid in effective decision making for cancer therapy.

With a video data source, such as multispectral video acquired during administration of fluorescent tracers, extraction of timeresolved data typically requires the compensation of motion. While this can be done manually, which is arduous, or using off-the-shelf object tracking software, which often yields unsatisfactory performance, we present an algorithm which is simple and performant. Most importantly, we provide an open-source implementation, with an easy-to-use interface for researchers not inclined to write their own code, as well as Python modules that can be used programmatically.

The wide availability of near infrared light sources in interventional medical imaging stacks enables non-invasive quantification of perfusion by using fluorescent dyes, typically Indocyanine Green (ICG). Due to their often leaky and chaotic vasculatures, intravenously administered ICG perfuses through cancerous tissues differently. We investigate here how a few characteristic values derived from the time series of fluorescence can be used in simple machine learning algorithms to distinguish benign lesions from cancers. These features capture the initial uptake of ICG in the colon, its peak fluorescence, and its early wash-out. By using simple, explainable algorithms we demonstrate, in clinical cases, that sensitivity (specificity) rates of over 95% (95%) for cancer classification can be achieved.

Reinforcement learning (RL) algorithms aim to learn optimal decisions in unknown environments through the experience of taking actions and observing the rewards gained. In some cases, the environment is not influenced by the actions of the RL agent, in which case the problem can be modeled as a contextual multi-armed bandit, and lightweight myopic algorithms can be employed. On the other hand, when the RL agent’s actions affect the environment, the problem must be modeled as a Markov decision process, and more complex RL algorithms are required, which take the future effects of actions into account. Moreover, in practice, it is often unknown from the outset whether or not the agent’s actions will impact the environment, and it is therefore not possible to determine which RL algorithm is most fitting. In this work, we propose to avoid this difficult decision entirely and incorporate a choice mechanism into our RL framework. Rather than assuming a specific problem structure, we use a probabilistic structure estimation procedure based on a likelihood-ratio (LR) test to make a more informed selection of the learning algorithm. We derive a sufficient condition under which myopic policies are optimal, present an LR test for this condition, and derive a bound on the regret of our framework. We provide examples of real-world scenarios where our framework is needed and provide extensive simulations to validate our approach.