Biology subjects

Leech, K. A.

Publications and source records attributed to Leech, K. A..

5 recordsLinked to original sources

bioRxiv2026

CALM-VLM: CALIBRATION AND SELECTIVE PREDICTION IN VISION-LANGUAGE MODELS FOR RELIABLE BRAIN MRI CLASSIFICATION

AO_SCPLOWBSTRACTC_SCPLOWRecent advances in vision-language models (VLMs) have demonstrated strong multimodal capabilities for medical image analysis. However, their confidence in diagnostic predictions is often unclear, limiting adoption in clinical settings. We introduce CALM-VLM (CALibration Mechanism for Vision-Language Models), which integrates confidence calibration and selective prediction into a generative 3D VLM. To create CALM-VLM, we fine-tuned the Med3DVLM architecture for Alzheimers disease (AD) and stroke classification as initial test cases. To improve reliability, we incorporated temperature scaling on the VLMs generative outputs. The calibrated model then selectively abstained from predictions when uncertain; this also improved its diagnostic accuracy. Experiments across multi-site MRI datasets, from 10 countries worldwide, show that CALM-VLM improved confidence relative to uncalibrated VLMs. Coverage-adjusted test receiver-operator characteristic curve-area under the curve (ROC-AUC) increased by 5% to 13% for both diagnostic tasks across independent test sets. Our calibrated VLM achieved a test ROC-AUC of 0.951 for AD classification and 0.905 for stroke classification. These findings highlight the importance of calibrated, uncertainty-aware VLMs for trustworthy neuroimaging AI.

neuroscience↗

bioRxiv2022

Gait speed and individual characteristics can be used to predict specific gait metric magnitudes in neurotypical adults.

BackgroundGait biofeedback is commonly used to reduce gait dysfunction in a variety of clinical conditions. In these studies, participants alter their walking to reach the desired magnitude of a specific gait parameter (the biofeedback target) with each step. Biofeedback of parameters such as anterior ground reaction force and step length have been well-studied. Yet, there is no standardized methodology to set the target magnitude of these parameters. Here we present an approach to predict the anterior ground reaction force and step length of neurotypical adults walking at different speeds as a potential method for personalized gait biofeedback. Research questionTo determine if anterior ground reaction force and step lengths achieved during neurotypical walking could be predicted using gait speed and participants demographic and anthropomorphic characteristics. MethodsWe analyzed kinetic and kinematic data from 51 neurotypical adults who walked on a treadmill at up to eight speeds. We calculated the average peak anterior ground reaction force and step length of the right lower extremity at each speed. We used linear mixed-effects models to evaluate the effect of speed, leg length, mass, and age on anterior ground reaction force and step length. We fit the model to data from 37 participants and validated predictions from the final models on an independent dataset from 23 participants. ResultsFinal prediction models for anterior ground reaction force and step length both included speed, speed squared, age, mass, and leg length. The models both showed strong agreement between predicted and actual values on an independent dataset. SignificanceAnterior ground reaction force and step length for neurotypical adults can be predicted given an individuals gait speed, age, leg length, and mass. This may provide a standardized method to personalize targets for individuals with gait dysfunction in future studies of gait biofeedback.

bioinformatics↗

bioRxiv2022

Speed-dependent biomechanical changes vary across individual gait metrics post-stroke relative to neurotypical adults

BackgroundGait training at fast speeds is recommended to reduce walking activity limitations post-stroke. Fast walking may also reduce gait kinematic impairments post-stroke. However, the magnitude of speed-dependent kinematic impairment reduction in people post-stroke relative to neurotypical adult walking patterns is unknown. ObjectiveTo determine the effect of faster walking speeds on gait kinematics post-stroke relative to neurotypical adults walking at similar speeds. MethodsWe performed a secondary analysis with data from 28 people post-stroke and 50 neurotypical adults treadmill walking at multiple speeds. We evaluated the effects of speed and group on individual spatiotemporal and kinematic metrics and performed k-means clustering with all metrics at self-selected and fast speeds ResultsPeople post-stroke decreased step length asymmetry and trailing limb angle impairment, reducing between-group differences at fast speeds. Speed-dependent changes in peak swing knee flexion, hip hiking, and temporal asymmetries exaggerated between-group differences. Our clustering analyses revealed two clusters. One represented neurotypical gait behavior, composed of neurotypical and post-stroke participants. The other characterized stroke gait behavior, comprised entirely of participants post-stroke. Cluster composition was largely consistent at both speeds, and the distance between clusters increased at fast speeds ConclusionsThe biomechanical effect of fast walking post-stroke varied across individual gait metrics. For participants within the stroke gait behavior cluster, speed-dependent changes did not lead to an overall gait pattern more similar to neurotypical adults. This suggests that combining fast walking with an approach to strategically target gait metrics with smaller speed-dependent changes may potentiate the biomechanical benefits of fast walking.

neuroscience↗

bioRxiv2020

People can learn new walking patterns without walking

Humans can learn many new walking patterns. People have learned to snowshoe up mountains, racewalk marathons, and march in precise synchrony. But what is required to learn a new walking pattern? Here, we demonstrate that people can learn new walking patterns without actually walking. Through a series of experiments, we observe that stepping with only one leg can facilitate learning of an entirely new walking pattern (i.e., split-belt treadmill walking). We find that the nervous system learns from the relative motion between the legs - whether or not both legs are moving - and can transfer this learning to novel gaits. We also show that locomotor learning requires active movement: observing another person adapt their gait did not result in significantly faster learning. These findings reveal that people can learn new walking patterns without bilateral gait training, as stepping with one leg can facilitate adaptive learning that transfers to novel gait patterns.

neuroscience↗

bioRxiv2019

Persons post-stroke restore step length symmetry by walking asymmetrically

BackgroundRestoration of step length symmetry is a common rehabilitation goal after stroke. Persons post-stroke often retain the capacity to walk with symmetric step lengths (\"symmetric steps\"); however, the resulting walking pattern remains effortful. Two key questions with direct implications for rehabilitation have emerged: 1) how do persons post-stroke generate symmetric steps, and 2) why do symmetric steps remain so effortful?\n\nObjectiveTo understand how persons post-stroke generate symmetric steps and how the resulting gait pattern relates to the metabolic cost of transport.\n\nMethodsTen persons post-stroke walked on an instrumented treadmill under two conditions: preferred walking and symmetric stepping (using visual feedback). We recorded kinematic, kinetic, and metabolic data during both conditions.\n\nResultsPersons post-stroke restored step length symmetry using energetically expensive, asymmetric patterns. Impaired paretic propulsion and abnormal vertical movement of the center of mass were evident during both preferred walking and symmetric stepping. These deficits contributed to diminished positive work performed by the paretic limb on the center of mass in both conditions. Decreased positive paretic work correlated with increased metabolic cost of transport, decreased self-selected walking speed, and increased asymmetry in limb kinematics.\n\nConclusionsIt is important to consider the mechanics used to restore symmetric steps when designing interventions to improve walking after stroke. Facilitating symmetric steps via increased paretic propulsion or enabling paretic limb advancement without excessive vertical movement may enable persons post-stroke to walk with a less effortful, more symmetric gait pattern.

neuroscience↗