Background: Dementia diagnosis is challenging and often delayed. Brain imaging techniques such as single-photon emission computed tomography (SPECT) imaging can help identify subtle changes in brain perfusion. Artificial intelligence methods may support results interpretation for early diagnosis. Objective: To develop and validate multivariate models for the early diagnosis of Alzheimer's disease (AD), using brain perfusion SPECT imaging and interpretable artificial intelligence methods in a real-world clinical setting. Methods: Two logistic regression models were developed using a training dataset of 420 SPECT scans and tested on an independent clinical dataset of 443 scans. Model 1 was designed to identify abnormal perfusion patterns, while Model 2 identified perfusion changes associated with AD. Input features were extracted from anatomical volumes of interest, with feature selection performed using the Minimum Redundancy Maximum Relevance (MRMR) algorithm.Results The models demonstrated good classification performance using real-world clinical data. Model 1 achieved an area under receiver operator characteristic (AUROC) Curve of 0.89 (Sensitivity 76%, Specificity 87%) in identifying abnormal brain perfusion. Model 2 achieved an AUROC of 0.86 (Sensitivity 87%, Specificity 72%) in identifying AD. Conclusions: Multivariate logistic regression models trained on real-world clinical data show promise as clinical decision support tools for the diagnosis of AD from brain perfusion SPECT imaging. The models use features from clinically relevant brain regions, which enhances interpretability. Future research should focus on expanding model applicability to other dementia types and on prospective evaluation of their utility in improving diagnostic accuracy, consistency, and care pathways in diverse clinical environments.

We report the results of a Phase I radiation dose escalation study using an yttrium-90 (90Y) labelled anti-CD66 monoclonal antibody given with standard conditioning regimen for patients receiving haematopoietic stem cell transplants for myeloid leukaemia or myeloma. The 90Y-labelled anti-CD66 was infused prior to standard conditioning. In total, 30 patients entered the trial and 29 received 90Y-labelled mAb, at infused radiation activity levels of 5, 10, 25, or 37.5 megaBequerel (MBq)/kg lean body weight. A prerequisite for receiving the 90Y-labelled mAb was favourable dosimetry determined by single-photon emission computerised tomography (SPECT) dosimetry following administration of indium-111 (111In) anti-CD66. Estimated absorbed radiation doses delivered to the red marrow demonstrated a linear relationship with the infused activity of 90Y-labelled mAb. At the highest activity level of 37.5 MBq/kg, mean estimated radiation doses for red marrow, liver, spleen, kidneys and lungs were 24.6 ± 5.6 Gy, 5.8 ± 2.7 Gy, 19.1 ± 8.0 Gy, 2.1 ± 1.1 and 2.2 ± 0.9, respectively. All patients engrafted, treatment-related mortality 1-year post-transplant was zero. Toxicities were no greater than those anticipated for similar conditioning regimens without targeted radiation. The ability to substantially intensify conditioning prior to haematopoietic stem cell transplantation without increasing toxicity warrants further testing to determine efficacy. clinicaltrials.gov identifier: NCT01521611.

Objective Predictive algorithms trained from historical data and deployed in dynamic environments are at risk from data drift. Machine learning models using data collected by sensors making continuous measurements could be impacted by both changes in the device itself and their users, driving drift and impacting safety. To maintain predictive performance, algorithms must be continuously monitored and tuned to overcome fundamental changes to both input data (covariate shift) and the relationship with the output (concept drift). Here, we aim to understand how changes to user behaviour, physiology and sensors could impact the safety of models using automated sensor readings from continuous glucose monitors (CGM).Methods and analysis In this paper, we investigate how data drift in a machine learning model trained to predict short-term risk from blood glucose control for individuals with type-1 diabetes. We simulate how changes in both user behaviour and accuracy of the sensor could lead to covariate shift and concept drift. For each scenario, we quantify the changes to input data (Jensen-Shannon divergence), the impact to model performance metrics and the explainability of the model (ie, shift in feature importance).Results We demonstrate that using a combination of covariate shift detection, multiple performance metrics and feature importance offers a powerful methodology of identifying different types of drifts in sensor data. For blood glucose management, our scenarios focused on user behaviour (ie, changes to blood glucose dynamics and CGM use) and device/sensor noise and variability, finding more simplistic approaches to drift detection could incorrectly identify risk to model safety.Conclusion Machine learning and AI can enhance clinical decision-making, but often lack the transparency required to ensure ongoing safety. Combining complementary monitoring techniques enables clearer identification of changes in data or model behaviour, helping determine when retraining or intervention is needed.

Introduction:Disease-modifying treatments such as monoclonal antibodies can be highly effective in chronic inflammatory diseases such as COPD, but often fail in clinical trials due to heterogeneity of inflammation and imperfect tools to stratify patients to select optimal therapeutic approaches. Molecular imaging provides the potential to transform precision medicine in this field. Methods:We developed and tested a novel molecular imaging platform using therapeutic monoclonal antibodies labelled with SPECT-CT detectable markers to quantify in vivo tumour necrosis factor (TNF) involved in chronic lung inflammation in humans. We undertook a proof-of-concept clinical study involving participants with COPD and healthy controls. Participants underwent SPECT-CT imaging at 6- and 24 h following injection of 99mTc-anti-TNF. Segmentation of lung regions and 99mTc-anti-TNF activity quantification was undertaken using novel semi-automated and AI-driven approaches. Results:A significant increase in normalised activity, representing increased TNF inflammatory activity, was seen between the two time-points in the COPD group (mean±sd: 64.88±31.04%, p=0.029) and not in healthy controls (35.38±34.33%, p=0.110). However, analysis at a single time-point revealed higher normalised activity in the healthy group. We demonstrated that pulmonary blood vessel density and degree of emphysema were strongly correlated with this activity signal and identified as confounding factors, highlighting the need to address differences in target-organ characteristics in COPD. Experimental methods to adjust for these factors were developed for organ-specific signal quantification. Conclusions:We report novel analysis techniques for molecular imaging of the human lung, presenting a platform which provides new insights into complex inflammatory disease and future precision medicine approaches.

Reinforcement learning (RL) holds promise for supporting personalised decision-making in healthcare, but existing approaches often struggle to incorporate patient expertise and individual preferences, key components for clinically viable AI systems. This work introduces PAINT (Preference Adaptation for Individualised Treatment), a general framework for preference-guided offline RL in safety-critical settings. PAINT combines sketch-based reward annotation with safety-constrained policy optimisation, enabling fine-grained preference capture from historical patient data without requiring action labels. A reward model trained on this feedback guides offline RL while supporting tunable sensitivity to preference signals and enforcing clinical safety constraints. Using type 1 diabetes (T1D) management as a case study, in-silico evaluation with the FDA-accepted T1D simulator demonstrates that can PAINT reduces patient risk by 15% over commercial baselines under guidance, while enabling preference-driven adaptations such as improved management during challenging mealtime events and enhanced robustness to dosing errors. The method further shows resilience to real-world challenges including sample size, annotation noise, and inter-patient variability. These findings suggest PAINT offers a practical pathway for integrating human feedback into offline RL in patient settings, with broader implications for developing trustworthy and adaptive AI systems in healthcare.

BACKGROUND:The occurrences of acute complications arising from hypoglycemia and hyperglycemia peak as young adults with type 1 diabetes (T1D) take control of their own care. Continuous glucose monitoring (CGM) devices provide real-time glucose readings enabling users to manage their control proactively. Machine learning algorithms can use CGM data to make ahead-of-time risk predictions and provide insight into an individual's longer term control. METHODS:We introduce explainable machine learning to make predictions of hypoglycemia (<70 mg/dL) and hyperglycemia (>270 mg/dL) up to 60 minutes ahead of time. We train our models using CGM data from 153 people living with T1D in the CITY (CGM Intervention in Teens and Young Adults With Type 1 Diabetes)survey totaling more than 28 000 days of usage, which we summarize into (short-term, medium-term, and long-term) glucose control features along with demographic information. We use machine learning explanations (SHAP [SHapley Additive exPlanations]) to identify which features have been most important in predicting risk per user. RESULTS:Machine learning models (XGBoost) show excellent performance at predicting hypoglycemia (area under the receiver operating curve [AUROC]: 0.998, average precision: 0.953) and hyperglycemia (AUROC: 0.989, average precision: 0.931) in comparison with a baseline heuristic and logistic regression model. CONCLUSIONS:Maximizing model performance for glucose risk prediction and management is crucial to reduce the burden of alarm fatigue on CGM users. Machine learning enables more precise and timely predictions in comparison with baseline models. SHAP helps identify what about a CGM user's glucose control has led to predictions of risk which can be used to reduce their long-term risk of complications.

Blood glucose simulation allows the effectiveness of type 1 diabetes (T1D) management strategies to be evaluated without patient harm. Deep learning algorithms provide a promising avenue for extending simulator capabilities; however, these algorithms are limited in that they do not necessarily learn physiologically correct glucose dynamics and can learn incorrect and potentially dangerous relationships from confounders in training data. This is likely to be more important in real-world scenarios, as data is not collected under strict research protocol. This work explores the implications of using deep learning algorithms trained on real-world data to model glucose dynamics. Free-living data was processed from the OpenAPS Data Commons and supplemented with patient-reported tags of challenging diabetes events, constituting one of the most detailed real-world T1D datasets. This dataset was used to train and evaluate state-of-the-art glucose simulators, comparing their prediction error across safety critical scenarios and assessing the physiological appropriateness of the learned dynamics using Shapley Additive Explanations (SHAP). While deep learning prediction accuracy surpassed the widely-used mathematical simulator approach, the model deteriorated in safety critical scenarios and struggled to leverage self-reported meal and exercise information. SHAP value analysis also indicated the model had fundamentally confused the roles of insulin and carbohydrates, which is one of the most basic T1D management principles. This work highlights the importance of considering physiological appropriateness when using deep learning to model real-world systems in T1D and healthcare more broadly, and provides recommendations for building models that are robust to real-world data constraints.

Engaging end user groups with machine learning (ML) models can help align the design of predictive systems with people's needs and expectations. We present a co-design study investigating the benefits and challenges of using computational notebooks to inform ML models with end user groups. We used a computational notebook to engage young adults, carers, and clinicians with an example ML model that predicted health risk in diabetes care. Through co-design workshops and retrospective interviews, we found that participants particularly valued using the interactive data visualisations of the computational notebook to scaffold multidisciplinary learning, anticipate benefits and harms of the example ML model, and create fictional feature importance plots to highlight care needs. Participants also reported challenges, from running code cells to managing information asymmetries and power imbalances. We discuss the potential of leveraging computational notebooks as interactive co-design tools to meet end user needs early in ML model lifecycles.

Aim Brain SPECT can detect early changes in perfusion to support the diagnosis of dementia. Amyloid Beta (Aβ) plaque deposits and aggregation of hyperphosphorylated tau in neurofibrillary tangles are hallmark pathologies of Alzheimer disease (AD). The aim of this study was to identify changes in brain perfusion patterns (neuroimaging signatures) with CSF amyloid and tau. Materials and Methods 91 participants’ SPECT scans and CSF samples from a heterogenous clinical cohort were analysed. AD biomarkers including Aβ42 and pTau were measured using a chemiluminescent enzyme immunoassay. Statistical Parametric Mapping (SPM) was used to quantify brain perfusion diffe- rences in SPECT scans in comparison to a database of healthy controls. Specifically, whole brain analysis was performed in SPM with Aβ42 and pTau added as covariates to univariate linear regression models Results Decreasing Aβ42 corresponds to significant reduction in perfusion in the parietal and temporal lobes, bilaterally (cluster centres in angular gyrus and medial temporal areas). Increasing pTau corresponds to reduction in brain perfusion medially in the parietal lobe (cluster centre in precuneus). Conclusion CSF Aβ42 and pTau showed distinct neuroimaging signatures in brain which could be pre- dictive of regional brain injury.

Neuroinflammation and activation of the immune system is an integral part of Alzheimer’s Dementia (AD) pathology. Inflammatory mediators exacerbate the production of amyloid-β, the propagation of tau pathology and neuronal loss. This study evaluates whether CSF markers of inflammation can help evaluate changes in amyloid and tau pathology in a heterogenous clinical population. CSF samples from 105 patients referred to the Wessex Neurology Clinic due to cognitive complaints were analysed. Measurements of AD biomarkers were used to classify the samples based on previously published thresholds (Ab42<680pg/ml for an amyloid positive test, pTau>56pg/ml for a Tau positive test and Total Tau>355pg/ml for a test positive for neurodegeneration). Based on these biomarker results, the likelihood of AD was evaluated using the Paris Lille Montpellier (PLM) scale. 102 markers of inflammation were measured on CSF using the Mesoscale and OLINK platforms. Receiver Operator Characteristic (ROC) curves were used to evaluate if inflammation markers can accurately identify patients with amyloid and tau pathology. 56 patients were amyloid positive, 43 were tau positive and 44 were positive for neurodegeneration. 52 different inflammation markers were detected in over 90% of samples. From these, 26 markers correlate significantly with pTau and Total Tau measurements, while no markers correlate with Ab42 measurements. Adenosine Deaminase (ADA), an enzyme of purine metabolism and marker of cellular immunity, most strongly correlates with pTau and Total Tau (Spearman’s Rho 0.62 and 0.60 respectively, p<0.001). Analysis of Variance indicates significantly higher levels of ADA in patients with higher PLM scores and thus higher likelihood of AD (p<0.001). Finally, ROC analysis indicates that ADA can identify Tau positive patients (pTau>56pg/ml) with an Area Under the Curve (AUC) of 0.76 and Neurodegeneration positive patients (Total Tau>355pg/ml) with an AUC of 0.82. In this clinical patient cohort, inflammation levels increase with tau pathology but do not change with amyloid. ADA, a marker of cellular immunity, provides a sensitive and specific marker of Tau and Neurodegeneration in AD. Further work is required to assess the prognostic capabilities of ADA in larger patient cohorts and when used in conjunction with established AD biomarkers.

Introduction: COPD is a heterogenous disease. Molecular imaging of inflammation could define endotypes, but heterogeneity of lung structure complicates this process. We aimed to develop methods to quantify inflammatory cytokines in heterogenous lung tissue as potential targets for therapy. Methods: Using SPECT-CT imaging, we developed techniques to quantify cytokine activity. Five patients with COPD and 5 healthy volunteers were recruited under ethically approved informed consent. They underwent SPECT-CT of the lungs at 6 (+/- 1) and 24 (+/- 4) hours after infusion of 99mTc-anti-TNF-α to quantify TNF-α activity in the lungs. Quantification was normalised to aortic arch signal to account for biological clearance. Results: Median normalised SPECT counts (CN) were higher in the healthy group at both time points. Strong correlations were seen between CN and both blood vessel density and emphysema quantification (figure 1). A regression model to correct for emphysema revealed higher CN at both time points in the COPD group, but differences were not statistically significant. Conclusions: Molecular imaging of inflammatory cytokines is affected by key confounding factors, and analysis techniques should account for structural heterogeneity.

Introduction: Multiple inflammatory endotypes exist as targets for therapy in COPD but novel precision medicine approaches are needed to define these. We aim to develop non-invasive imaging methods to quantify specific inflammatory cytokines in the lung as potential targets for therapy. Methods: Using SPECT-CT, we developed techniques to quantify cytokine activity. Five patients with COPD and 5 healthy volunteers were recruited and gave informed consent. They underwent SPECT-CT of the lungs at 6 (+/- 1) and 24 (+/- 4) hours after infusion of 99mTc-anti-TNF-α to quantify TNF-α activity in the lungs. Quantification (figure 1) was normalised to aortic arch signal to account for biological clearance. Results: Isotope signals were quantified at 6 and 24 hours. Median normalised lung SPECT counts (CN) were higher in the healthy group at both time points. However, the increase in CN at 24 hours calculated as a percentage of the 6-hour scan (a measure of tissue bound signal) was higher in the COPD group at 64.88% +/- (SD) 31.04 compared with 35.38% +/- 34.33 in the healthy group - a significant change in the COPD group (p=0.029, paired t-test) but not the healthy. Conclusions: This proof-of-concept study provides early evidence that molecular imaging of inflammatory cytokines is possible in the lungs of patients with COPD.

BACKGROUND:Single photon emission tomography (SPECT) can detect early changes in brain perfusion to support the diagnosis of dementia. Inflammation is a driver for dementia progression and measures of inflammation may further support dementia diagnosis.OBJECTIVE:In this study, we assessed whether combining imaging with markers of inflammation improves prediction of the likelihood of Alzheimer's disease (AD).METHODS:We analyzed 91 participants datasets (Institutional Ethics Approval 20/NW/0222). AD biomarkers and markers of inflammation were measured in cerebrospinal fluid. Statistical parametric mapping was used to quantify brain perfusion differences in perfusion SPECT images. Logistic regression models were trained to evaluate the ability of imaging and inflammation markers, both individually and combined, to predict AD.RESULTS:Regional perfusion reduction in the precuneus and medial temporal regions predicted Aβ42 status. Increase in inflammation markers predicted tau and neurodegeneration. Matrix metalloproteneinase-10, a marker of blood-brain barrier regulation, was associated with perfusion reduction in the right temporal lobe. Adenosine deaminase, an enzyme involved in sleep homeostasis and inflammation, was the strongest predictor of neurodegeneration with an odds ratio of 10.3. The area under the receiver operator characteristic curve for the logistic regression model was 0.76 for imaging and 0.76 for inflammation. Combining inflammation and imaging markers yielded an area under the curve of 0.85.CONCLUSIONS:Study results showed that markers of brain perfusion imaging and markers of inflammation provide complementary information in AD evaluation. Inflammation markers better predict tau status while perfusion imaging measures represent amyloid status. Combining imaging and inflammation improves AD prediction.

Blood brain barrier dysfunction amplifies neuroinflammation, which may drive Alzheimer’s Dementia (AD) pathology. Regional cerebral blood flow (RCBF), measured by HMPAO SPECT, is an established biomarker for AD diagnosis. Matrix Metalloproteinase-10 (MMP-10), an enzyme involved in blood brain barrier function through regulating the breakdown of extracellular matrix, was recently proposed as a biomarker of progression to AD. In this study, we examine the relationship between RCBF and levels of MMP-10 in the cerebrospinal fluid (CSF). Datasets of 91 participants from a heterogenous clinical cohort, investigated for dementia due to cognitive complaints were analysed. CSF levels of MMP-10 were measured using the OLINK proximity extension array platform. HMPAO SPECT scans were analysed using Statistical Parametric Mapping (SPM). A univariate linear regression model was used in SPM to quantify the impact of MMP-10 changes on brain perfusion. SPM results showed that higher levels of MMP-10 in CSF are associated with significant reduction in RCBF (family-wise error corrected p<0.05). The neuroimaging signature of changes in RCBF with increasing MMP-10 is outlined in Figure 1. The main cluster of reduction in perfusion is on the right temporal lobe with an additional cluster on the right medial frontal lobe. Increased levels of MMP-10 in CSF have been associated with blood brain barrier vulnerability and faster cognitive decline in AD. Here we identified a right sided neuroimaging signature in RCBF with increasing levels MMP-10. This may indicate that by the time of symptoms onset the right side is the fastest progressing as it is catching up with the left side. Right sided changes have been previously associated with delusions, disinhibition and irritability in AD and linked with increased carer burden. The significant reduction in RCBF of the right temporal lobe identified in our study, further reinforces a role for MMP-10 as a marker of progression to AD. Acknowledgement: Dr Sofia Michopoulou, is funded through an Integrated Clinical Academic Lectureship by Health Education England (HEE) / NIHR for this research project (NIHR301287). The views expressed in this publication are those of the authors.

The widespread adoption of effective hybrid closed loop systems would represent an important milestone of care for people living with type 1 diabetes (T1D). These devices typically utilise simple control algorithms to select the optimal insulin dose for maintaining blood glucose levels within a healthy range. Online reinforcement learning (RL) has been utilised as a method for further enhancing glucose control in these devices. Previous approaches have been shown to reduce patient risk and improve time spent in the target range when compared to classical control algorithms, but are prone to instability in the learning process, often resulting in the selection of unsafe actions. This work presents an evaluation of offline RL for developing effective dosing policies without the need for potentially dangerous patient interaction during training. This paper examines the utility of BCQ, CQL and TD3-BC in managing the blood glucose of the 30 virtual patients available within the FDA-approved UVA/Padova glucose dynamics simulator. When trained on less than a tenth of the total training samples required by online RL to achieve stable performance, this work shows that offline RL can significantly increase time in the healthy blood glucose range from 61.6±0.3% to 65.3±0.5% when compared to the strongest state-of-art baseline (p<0.001). This is achieved without any associated increase in low blood glucose events. Offline RL is also shown to be able to correct for common and challenging control scenarios such as incorrect bolus dosing, irregular meal timings and compression errors. The code for this work is available at: https://github.com/hemerson1/offline-glucose.

This work applies a back-projection approach to the reconstruction of images obtained with a 3D near-field coded aperture camera, a method derived from that used to analyse astronomical data collected by the INTEGRAL/IBIS instrument.1 The method used in this paper is a form of deconvolution that updates the image as the camera or the object under observation is moving, and as such could be applied to dynamic studies or as a real-time surgical probe. Multi-isotope parathyroid imaging has been identified as one of the optimal applications of this technique, where high sensitivity and high energy resolution would permit the differentiation between thyroid and parathyroid tissues on the surgical table, and aid in different types of thyroid surgery. The back-projection technique is combined with a 3D version of the iterative CLEAN algorithm, which reduces the effect of the systematic noise intrinsic to imaging with a non-perfect aperture.

Background: The occurrences of acute complications arising from hypoglycaemia and hyperglycaemia peak as young adults with type 1 diabetes (T1D) take control of their own care. Continuous glucose monitoring (CGM) devices provide real-time blood glucose readings enabling users to manage their control pro-actively. Machine learning algorithms can use CGM data to make ahead-of-time risk predictions and provide insight into an individual's longer-term control. Methods: We introduce explainable machine learning to make predictions of hypoglycaemia (< 70 mg/dL) and hyperglycaemia (> 270 mg/dL) 60 minutes ahead-of-time. We train our models using CGM data from 153 people living with T1D in the CITY survey totalling over 28000 days of usage, which we summarise into (short-term, medium-term, and long-term) blood glucose features along with demographic information. We use machine learning explanations (SHAP) to identify which features have been most important in predicting risk per user. Results: Machine learning models (XGBoost) show excellent performance at predicting hypoglycaemia (AUROC: 0.998) and hyperglycaemia (AUROC: 0.989) in comparison to a baseline heuristic and logistic regression model. Conclusions: Maximising model performance for blood glucose risk prediction and management is crucial to reduce the burden of alarm-fatigue on CGM users. Machine learning enables more precise and timely predictions in comparison to baseline models. SHAP helps identify what about a CGM user's blood glucose control has led to predictions of risk which can be used to reduce their long-term risk of complications.