Nils Daniel Forkert

PurposeAlzheimer’s disease (AD) is the leading cause of dementia, characterized by the accumulation of amyloid-β and tau aggregates in the brain. Current diagnostic approaches primarily detect mature plaques, but growing evidence suggests that pre-fibrillar aggregates are present earlier in normal-appearing tissue and may provide a valuable window for intervention. We propose the first proof-of-concept end-to-end deep learning (DL) model to identify AD signatures directly from normal-appearing tissue (i.e., background parenchyma) of raw spectral fluorescence data.ApproachTo develop and evaluate our DL model, we used one million signals extracted from spectral fluorescence images of brain cortex sections collected from wild-type and 5xFAD mice (a transgenic strain carrying five familial human AD mutations).ResultsOur results demonstrate that AD-related signatures can indeed be detected in normal-appearing brain cortex regions. Experiments revealed that model performance depends on input dimensions, training set size, and the preprocessing kernel size. The optimal configuration used 1000×32 input, 9000 training samples, and a 3×3 kernel, achieving ≥95% accuracy, sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve. Importantly, no mice were misclassified, and saliency analysis indicated that a broad range of spectral fluorescence wavelengths contributed to the model’s decisions.ConclusionsThese findings highlight the potential of DL-based analysis of spectral fluorescence data to enable earlier detection of AD-related changes, supporting a shift toward a more proactive strategy for biomarker discovery.

Multi-echo functional MRI (fMRI) improves signal quality by exploiting the fact that true blood-oxygenation-level-dependent (BOLD) signals decay with T2* across echo times while many artifacts do not. By decomposing multi-echo fMRI data using independent component analysis (ICA), each component’s signal decay across echo times can be used to classify it as BOLD-like or non-BOLD-like. The tedana software classifies ICA components using metrics that capture this differential decay via rule-based decision trees. However, approximately 66% of components in our cohort were not classified with high confidence using tedana, leaving residual artifacts that limit downstream analysis sensitivity. To address this gap, we developed a multimodal convolutional neural network (CNN) that jointly learns from the spatial maps and time series of ICA components of optimally combined resting-state multi-echo data, using training labels derived from these multi-echo decay metrics, to more accurately classify these components. Trained on 12,750 high-confidence components from 229 subjects with perfect class balance, the model achieved 98.6% accuracy on held-out data and flagged as artifacts 10,704 components (33.3% of the non-high-confidence set) that tedana had accepted. Critically, denoising with CNN classifications significantly improved network-based contrast-to-noise ratio (CNR) by 6.20% (p=0.011), with six of seven resting-state networks showing gains. Furthermore, CNN-based denoising increased network modularity (Louvain Q +6.8%, p=0.0006) and showed numerically higher CNR test-retest reliability (ICC: 0.664 vs. 0.622). Ablation studies showed that the CNN’s CNR advantage arises from retaining components that tedana rejects, while the additional components it rejects can be removed without loss of network-level signal. On two independent external datasets, the CNN generalized without retraining: CNR gains were directionally consistent but non-significant on a resting-state cohort (+2.6%, p=0.36), whereas on a more dissimilar task-fMRI cohort (four echoes, different scanner, TR = 1.615 s) it significantly improved network CNR over tedana (+16.7%, p<0.001) while performing comparably to tedana in tSNR and task activation, which was enhanced over the non-denoised baseline. SmoothGrad analysis showed broad cortical saliency consistent with resting-state networks, strongest at inferior boundaries, for accepted components, and lower-magnitude edge saliency for rejected ones, confirming neurobiologically meaningful feature learning. These findings demonstrate that learned spatiotemporal features provide superior multi-echo ICA classification compared to traditional metric-based decision trees, offering an interpretable, data-driven solution for multi-echo fMRI denoising.

Advances in scale and depth across omics, biomarker, and imaging technologies now resolve molecular, cellular, tissue, and system-level biology in unprecedented detail. These data create vast opportunities to represent biological state, predict biological response, and guide experimental intervention in medicine. However, contemporary data science and human capacity for interpretation fall short at this scale. Artificial intelligence has become a principal focus in attempting to overcome challenges in biological modeling. Frontier modelling approaches learn representations from heterogeneous, multimodal datasets, providing tools to organize, characterize, and attempt to predict biological states under defined contexts. However, the value of these models is thus far limited, with little evidence of translation. Progress beyond the contemporary depends on appropriate design of computation around the architecture of biology: distinguishing context-specific associations from relationships that remain predictive across biological conditions, encoding understanding of where observed correlations converge or diverge within mechanisms, and consolidating interactions across molecule, cell, tissue, and system. In this work, we examine in silico biology from a systems perspective: computational systems that learn representations of biological state, predict state transitions, prioritize experimentally testable hypotheses, and guide intervention in living systems. We organize the field around three intertwined problems. The first is the representation problem: Learning biological state from heterogeneous, sparse, disconnected measurements. The second is the causality problem: Predicting how biological states maintain and respond to stimuli. The third is the intervention problem: Validating computationally perceived phenomena. In our outlook, the next frontier is not defined by architecture, modality, or information scaleup alone. It will emerge through cohesive representation across contexts, domains, and scales; interrogation of causality against computational design within and beyond the data, and through iterative computational-experimental intervention to refine both data and architecture to fill the gaps that contextualize causality against biology broadly.

Perinatal stroke affects millions of individuals worldwide, often leading to lifelong complications for those who survive and who require targeted rehabilitation to limit disability. Although biological brain age prediction may be a valuable biomarker to analyze neurodevelopment following perinatal stroke and guide rehabilitation, there have been no prior studies exploring its utility. Therefore, in this work, we analyzed neurodevelopment in children with two forms of perinatal stroke, namely arterial ischemic stroke (AIS) and periventricular venous infarction (PVI), by using T1‐weighted neuroimaging data and machine learning‐based biological brain age prediction at a global and voxel level. Specifically, we analyzed trends in the brain age gap (BAG) at both global and voxel‐wise levels in the contralesional hemisphere, alongside correlation analyses with motor scores in stroke cohorts. Global and voxel‐wise biological brain age prediction machine learning models were developed and trained using 5969 T1‐weighted MRI scans of typically‐developing children (mean age: 12.11 ± 2.72 years). These trained models were then applied to T1‐weighted MRI data from N = 105 subjects with perinatal stroke (mean age: 11.41 ± 3.27 years) and N = 105 age‐ and sex‐matched controls. The Wilcoxon signed‐rank test and Mann–Whitney U‐test were used to identify differences in BAGs in the stroke vs. controls subgroups, and AIS vs. PVI subgroups, respectively. Spearman's correlation coefficient was used to identify relationships between BAGs and motor scores. Lastly, sex‐specific analyses were performed to identify sexually dimorphic characteristics in the data. Children with perinatal stroke exhibited, on average, more positive global and voxel‐wise BAGs, particularly those with AIS. Motor scores were negatively correlated with global and voxel‐wise BAGs in the contralesional hemisphere. The correlations between BAGs and motor scores differed between the sexes. This is the first study to propose and explore voxel‐wise brain age prediction for a pediatric cohort and the first to utilize brain age prediction to study perinatal stroke. Further development of these methods may reveal biomarkers that are valuable to study other pediatric diseases and promote personalized rehabilitation for affected individuals.

Background: Mild behavioral impairment (MBI), characterized by later-life emergence of persistent neuropsychiatric symptoms (NPS), is an early clinical indicator of dementia risk. Global MBI has been associated with Alzheimer's disease (AD) pathology; studies have also explored MBI domains. Prior work has linked MBI-apathy to AD cerebrospinal fluid (CSF) biomarkers, but whether associations are detectable using plasma-based biomarkers such as phosphorylated tau (p-tau) is unknown. Establishing such relationships is critical, as plasma biomarkers are more accessible than CSF. Objective: To explore cross-sectional and longitudinal associations between MBI-apathy and plasma p-tau ₁₈₁ levels using Alzheimer's Disease Neuroimaging Initiative data. Methods: Older adults with normal cognition or mild cognitive impairment were categorized as MBI-apathy (n = 69), non-MBI NPS (n = 112), and no-NPS (n = 215) based on Neuropsychiatric Inventory scores and symptom persistence over one year. Linear regression modelled cross-sectional associations between NPS group and plasma p-tau ₁₈₁ , adjusting for age, sex, education, apolipoprotein E4 status, and Mini-Mental State Examination score. Hierarchical linear mixed-effects modelling assessed associations over two and three years, including time-by-NPS group interactions. Results: MBI-apathy was associated with significantly higher plasma p-tau ₁₈₁ levels at baseline (24.05% [6.06–45.08%]; adjusted p = 0.014), and over two (26.46% [7.24–49.12%]; adjusted p = 0.012) and three years (29.28% [10.17–51.72%]; adjusted p = 0.004) compared to no-NPS. No significant associations were observed for non-MBI NPS. In sensitivity analyses, non-MBI apathy was not associated with plasma p-tau ₁₈₁ at baseline (−9.96% [−32.69–20.44%]; unadjusted p = 0.478). Conclusions: MBI-apathy is associated with elevated plasma p-tau ₁₈₁ cross-sectionally and longitudinally, supporting MBI-apathy as a potential proxy marker of tau pathology for early AD detection.

PurposeAssessing treatment response in bone metastases from non-small cell lung cancer (NSCLC) remains a major clinical challenge, particularly for patients receiving immune checkpoint inhibitors (ICIs). The existing response criteria are not optimized for osseous disease, leading to inconsistent evaluation. We aimed to develop and validate a radiomics-based machine learning (ML) framework to non-invasively distinguish immunotherapy response categories—progression, stable disease, and partial response—in NSCLC patients with bone metastases.ApproachChest computed tomography (CT) scans from 99 NSCLC patients were analyzed before and during ICI therapy. Bone structures were automatically segmented using TotalSegmentator, and 1051 radiomic features were extracted per time point. Clinical variables were incorporated as optional features. Three ML classifiers—random forest, XGBoost, and support vector machine—were trained using fivefold cross-validation. A multistep feature selection pipeline (correlation filtering, mutual information, recursive feature elimination, and ReliefF ranking) was applied. Model performance was evaluated using area under the curve (AUC), F1-score, accuracy, sensitivity, and specificity, with additional statistical testing using Kruskal–Wallis, Mann–Whitney U, bootstrapping, and permutation analysis.ResultsInter-rater agreement for radiological response categories was high (Cohen’s kappa = 0.91). Post-treatment radiomic features yielded the best performance. The random forest model achieved an AUC of 0.94, an F1-score of 0.79, an accuracy of 0.79, a sensitivity of 0.80, and a specificity of 0.83. Clinical features did not meaningfully improve performance. Models based on the largest lesion showed lower accuracy than those using the overall response.ConclusionsPost-treatment CT radiomics captured therapy-induced skeletal changes and enabled differentiation of immunotherapy response categories in NSCLC bone metastases. These findings highlight radiomics as a non-invasive tool for response assessment and guiding personalized treatment strategies.

Vision impairment remains a major global health challenge, and early diagnosis of retinal diseases is essential to prevent avoidable vision loss. Color fundus photography is widely used for retinal assessment, but manual interpretation is time-consuming and prone to inter-clinician variability. Although deep learning has improved automated retinal disease detection, most classification models lack transparency. We introduce RetCond, a self-explanatory generative classifier designed to provide both accurate predictions and counterfactual visual explanations. RetCond is a diffusion model repurposed as a classifier whose generative capabilities enable counterfactual synthesis for built-in explanations. These counterfactual images depict the retinal changes associated with alternative diagnoses, offering insight into the model’s decisions. We curated a diverse dataset of 19,565 images containing five retinal disease conditions to train and evaluate the model. Performance is assessed using standard classification metrics and by qualitatively and quantitatively evaluating the self-explanatory component. Comparisons are made against strong discriminative baselines. RetCond achieves a classification accuracy of 96.98%, comparable to RETFound (96.99%). In contrast to standard models, RetCond produces counterfactual images that highlight condition-specific retinal features, confirming its learned concepts and demonstrating its self-explanatory properties. These results show that RetCond matches state-of-the-art discriminative performance while addressing limitations of post-hoc interpretability techniques in standard deep learning models. RetCond, therefore, constitutes a contribution toward more trustworthy AI solutions in ophthalmology.

Background: Mild behavioural impairment (MBI), characterized by later-life emergence of persistent neuropsychiatric symptoms (NPS), is an early clinical indicator of dementia risk. MBI as a global construct has been associated with Alzheimer disease (AD) pathology; studies have also explored MBI domains. Prior work has linked MBI-apathy to cerebrospinal fluid (CSF) biomarkers of AD, but whether similar associations are detectable using plasma-based biomarkers such as phosphorylated tau (p-tau) is unknown. Establishing such relationships is critical, as plasma biomarkers are more accessible than CSF. Objective: To explore cross-sectional and longitudinal associations between MBI-apathy and plasma p-tau ₁₈₁ using Alzheimer’s Disease Neuroimaging Initiative data. Methods: Older adults with normal cognition or mild cognitive impairment were categorized as MBI-apathy (n=69), non-MBI NPS (n=112), and no-NPS (n=215) based on Neuropsychiatric Inventory scores and symptom persistence over one year. Linear regression modelled cross-sectional associations between NPS group and plasma p-tau ₁₈₁ levels, adjusting for age, sex, education, apolipoprotein E4 status, and Mini-Mental State Examination score. Hierarchical linear mixed-effects modelling assessed associations over two and three years, including time-by-NPS group interactions. Results: MBI-apathy was associated with significantly higher plasma p-tau ₁₈₁ levels at baseline (24.05% [6.06–45.08%]; adjusted p =0.014), and over two (26.46% [7.24–49.12%]; adjusted p =0.012) and three years (29.28% [10.17–51.72%]; adjusted p =0.004) compared to no-NPS. No significant associations were observed for non-MBI NPS. Conclusions: MBI-apathy is associated with elevated plasma p-tau ₁₈₁ cross-sectionally and longitudinally. These findings support MBI-apathy as a potential proxy marker of tau pathology for early AD detection.

Mild Behavioural Impairment (MBI) is defined by later-life onset of persistent behavioural changes and is recognized as a risk marker for cognitive decline and dementia. Apathy, a core MBI domain characterized by diminished interest, initiative, and emotional reactivity, can emerge before dementia and is hypothesized to be associated with structural brain changes. While previous studies have explored Alzheimer disease (AD)-related neuroanatomical substrates of apathy in the dementia clinical stage, few have investigated these associations in cognitively normal (CN) or mild cognitive impairment (MCI) individuals with persistent apathy consistent with MBI. Thus, this study explores structural brain differences between individuals with MBI-apathy and those without neuropsychiatric symptoms (no-NPS). Participants (n = 446; mean age = 69.6 years; 79.8% CN; 62.8% female) were drawn from the National Alzheimer’s Coordinating Center and categorized into MBI-apathy (n = 59) and no-NPS (n = 387) groups. Linear regressions were used to model associations between NPS group and regional brain measures, with adjustments for age, sex, years of education, apolipoprotein E4 carrier status, intracranial volume, and Mini-Mental State Examination score, with false discovery rate (FDR) correction for multiple comparisons. Primary outcomes included two predefined AD meta-regions-of-interest (ROIs): 1) thickness: a composite measure of mean cortical thickness across the entorhinal cortex, inferior temporal gyrus, middle temporal gyrus, inferior parietal lobule, fusiform gyrus, and precuneus; and 2) volume: a composite measure of mean cortical and subcortical grey matter volume across the hippocampus, entorhinal cortex, amygdala, middle temporal gyrus, inferior parietal lobule, and precuneus. Primary outcomes also included cortical thickness and grey matter volume among individual ROIs including the ventral striatum (VS), anterior cingulate cortex (ACC), orbitofrontal cortex (OFC), ventrolateral prefrontal cortex (vlPFC), and dorsolateral prefrontal cortex (dlPFC). MBI-apathy status was associated with significantly lower AD-meta-ROI cortical thickness ( Z -score difference [95% CI]; FDR-corrected p -value, -0.43 [-0.73 – [-0.12]]; 0.025) and lower AD meta-ROI grey matter volume (-0.50 [-0.71 – [-0.30]]; <0.001). MBI-apathy was also associated with significantly lower dlPFC thickness (-0.40, [-0.70 – [-0.09]]; 0.02) and volume (-0.28 [-0.50– [-0.06]]; 0.026) and lower OFC volume (-0.32, [-0.57 – [-0.07]]; 0.026) compared to the no-NPS group. Within a non-dementia sample, MBI-apathy was more strongly associated with established AD-vulnerable regions than with regions that have been traditionally implicated in apathy in dementia. Results suggests that during CN and MCI stages, MBI-apathy may reflect early AD-related neurodegeneration, with conventional apathy-related structural changes becoming more prominent as disease progresses.

Artificial intelligence (AI) has vast potential for improving healthcare delivery, but concerns regarding biases in these systems have raised important questions regarding fairness when deployed clinically. Most prior studies on fairness in clinical AI focus solely on performance disparities between subpopulations, which often fall short of connecting the technical outputs of AI systems with sociotechnical outcomes. In this work, we present a simulation-based approach to explore how statistical definitions of algorithmic fairness translate to fairness in long-term outcomes, using AI-assisted breast cancer screening as a case example. We evaluate four fairness criteria and their impact on mortality rates and socioeconomic disparities, while also considering how clinical decision makers’ reliance on AI and patients’ access to healthcare affect outcomes. Our results highlight how algorithmic fairness does not directly translate into fair and equitable outcomes, underscoring the importance of integrating sociotechnical perspectives to gain a holistic understanding of fairness in healthcare AI.