Methods and systems for multi-compartment body composition analysis

Abstract

A system and method for multi-compartment body composition analysis through computer vision and machine learning is provided. The system processes standardized digital images captured via mobile device cameras to generate body composition assessments including bone mineral content, skeletal muscle mass, fat mass, visceral adipose tissue, and total body water. Machine learning models are trained on paired datasets comprising digital images temporally aligned with measurements from medical imaging devices. The training methodology incorporates physiological mass-balance constraints that require predicted compartment values to satisfy interdependency relationships. Through direct inference from digital images without intermediate three-dimensional reconstruction, body volume computation, or anatomical keypoint detection, the system may achieve accuracy validated against medical imaging device measurements for accessible body composition monitoring. The technology enables longitudinal health tracking and diagnostic screening for conditions including osteoporosis, sarcopenia, and metabolic disorders using commodity smartphone hardware.

Description

METHODS AND SYSTEMS FOR MULTI-COMPARTMENT BODY COMPOSITION ANALYSIS INCORPORATION BY REFERENCE OF RELATED APPLICATIONS

[0001] This application claims the benefit under 35 U.S.C. § 119(e) of U.S. Provisional Patent Application No. 63 / 755,717, filed February 7, 2025, the entire contents of which are incorporated by reference herein and are made a part of this specification.BACKGROUND

[0002] Accurate assessment of body composition is foundational to modem healthcare, fitness, and wellness management. Quantitative measurement of bone density, muscle mass, fat distribution, and water composition enables screening and diagnosis of chronic conditions including osteoporosis, sarcopenia, cardiovascular diseases, and metabolic disorders. As prevalence of these conditions continues to increase globally, demand for accessible body composition monitoring has grown correspondingly.

[0003] Existing diagnostic solutions using specialized medical imaging devices face implementation barriers that limit widespread adoption and routine monitoring. These banners include infrastructure requirements, facility access limitations, cost considerations, and in some cases radiation exposure concerns. These limitations have motivated development of alternative approaches, though existing alternatives have not achieved the combination of accessibility, accuracy, and clinical utility required for widespread adoption.SUMMARY

[0004] Example embodiments described herein have innovative features, no single one of which is indispensable or solely responsible for their desirable attributes. Without limiting the scope of the claims, some of the advantageous features will now be summarized.

[0005] In some aspects, the techniques described herein relate to a computer-implemented method for generating a predictive model for body composition analysis, the method performed by one or more processors executing instructions stored in non-transitory computer-readable memory, the method including: for each of a plurality of training subjects, capturing a set of digital images using a mobile computing device, wherein each set includes a plurality of views of the subject in a standardized anatomical position; introducing controlled environmentalvariability across the captured image sets, wherein the environmental variability includes variation in at least two of: image background, ambient lighting conditions, device orientation angle, shadow characteristics, and subject clothing; for each training subject, acquiring body composition reference measurements from at least one medical imaging device selected from the group consisting of: Dual-Energy X-ray Absorptiometry (DEXA), Bioelectrical Impedance Analysis (BIA), Air Displacement Plethysmography (ADP), Magnetic Resonance Imaging (MRI), and Computed Tomography (CT) within a temporal window relative to the image capture, wherein the reference measurements are used for model training and are not required during subsequent inference on new subjects; associating the captured image sets with the corresponding body composition reference measurements and demographic data to form paired training data; preprocessing the captured images to isolate each subject from the image background; and training a computational model on the paired training data using backpropagation to iteratively adjust model parameters by minimizing a composite loss function, wherein the composite loss function includes: a prediction error term measuring deviation between model outputs and reference measurements, and; a physiological constraint term that penalizes violations of mass-balance relationships requiring predicted compartment masses to sum to the subject's measured body weight within a predefined tolerance; wherein the trained computational model performs direct inference from two-dimensional image pixel data to body composition outputs without computing intermediate three-dimensional reconstructions, body volume, or surface-geometry-based anthropometric measurements.

[0006] In some aspects, the techniques described herein relate to a method, wherein the physiological mass-balance relationships require that inferred compartment values sum to a value within a predefined tolerance of the subject’s measured body weight.

[0007] In some aspects, the techniques described herein relate to a method, wherein the plurality of views includes an anterior view, a posterior view, a left lateral view, and a right lateral view.

[0008] In some aspects, the techniques described herein relate to a method, wherein the at least one medical imaging device includes one or more devices capable of generating body composition reference measurements, including devices selected from the group consisting of: Dual-Energy X-ray Absorptiometry (DEXA), Bioelectrical Impedance Analysis (BIA), AirDisplacement Plethysmography (ADP), Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and multi-frequency bioimpedance spectroscopy systems.

[0009] In some aspects, the techniques described herein relate to a method, wherein the at least three physiologically distinct body composition compartments include: fat mass, lean mass, and at least one additional compartment selected from the group consisting of: bone mineral content, total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

[0010] In some aspects, the techniques described herein relate to a method, wherein the body composition compartments further include one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

[0011] In some aspects, the techniques described herein relate to a method, wherein the demographic data includes at least two of: age, biological sex, height, weight, and ethnicity.

[0012] In some aspects, the techniques described herein relate to a method, wherein training the computational model further includes training on longitudinal paired data including multiple image sets and corresponding reference measurements for individual training subjects captured at different time points.

[0013] In some aspects, the techniques described herein relate to a method, wherein the different time points are separated by intervals of approximately four to eight weeks.

[0014] In some aspects, the techniques described herein relate to a method, wherein training the computational model includes using a loss function including a term that penalizes deviations between the sum 1 to 10, wherein the threshold temporal window is ten minutes or less.

[0015] In some aspects, the techniques described herein relate to a computer-implemented method for determining body composition of a subject, the method performed by a server system including one or more processors and memory storing a trained computational model, the method including: receiving, from a mobile computing device, a set of digital images of the subject including a plurality of views of the subject; preprocessing the digital images to isolate the subject from the image background; applying a trained computational model to the preprocessed image data, wherein the trained computational model performs joint inference of body composition metrics through pixel-level analysis of two-dimensional images, generating predictions directly from image features; and outputting, as a result of the joint inference, predicted values for a plurality of body composition metrics including at least three physiologically distinct body composition compartments, wherein the predicted values are interdependent such that thevalue for any one compartment constrains the values for remaining compartments to satisfy physiological consistency relationships.

[0016] In some aspects, the techniques described herein relate to a method, wherein the trained computational model was trained on paired data including digital images and temporally aligned body composition reference measurements from at least one medical imaging device.

[0017] In some aspects, the techniques described herein relate to a method, wherein the temporal alignment includes image capture within a threshold temporal window of reference measurement acquisition.

[0018] In some aspects, the techniques described herein relate to a method, wherein the threshold temporal window is ten minutes or less.

[0019] In some aspects, the techniques described herein relate to a method, wherein the trained computational model was trained using a loss function that penalizes violations of physiological mass-balance constraints.

[0020] In some aspects, the techniques described herein relate to a method, wherein the plurality of views includes an anterior view, a posterior view, and at least one lateral view.

[0021] In some aspects, the techniques described herein relate to a method, wherein the at least three physiologically distinct body composition compartments include: fat mass, lean mass, and at least one additional compartment selected from the group consisting of: bone mineral content, total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

[0022] In some aspects, the techniques described herein relate to a method, further including outputting regional body composition measurements for body segments selected from the group consisting of: right arm, left arm, right leg, left leg, and trunk.

[0023] In some aspects, the techniques described herein relate to a method, further including receiving the subject's measured body weight, wherein the predicted compartment values are computed as percentages of the measured body weight to satisfy the physiological consistency relationships.

[0024] In some aspects, the techniques described herein relate to a method, wherein the predicted values for fat mass demonstrate a mean absolute error of less than five percent when evaluated against measurements from Dual-Energy X-ray Absorptiometry (DEXA).

[0025] In some aspects, the techniques described herein relate to a method, wherein the predicted values for fat mass demonstrate a mean absolute error of less than seven percent when evaluated against measurements from at least one medical imaging device.

[0026] In some aspects, the techniques described herein relate to a method, wherein the mean absolute error for fat mass is less than four percent when validated against DEXA measurements.

[0027] In some aspects, the techniques described herein relate to a method, wherein the trained computational model achieves prediction accuracy for body fat percentage with coefficient of determination (R2) greater than 0.65 when validated against reference measurements.

[0028] In some aspects, the techniques described herein relate to a method, wherein the coefficient of determination may be greater than 0.75 for measurements obtained within ninety days of reference device measurement.

[0029] In some aspects, the techniques described herein relate to a method, wherein the trained computational model was trained on a dataset including at least five hundred paired samples of digital images and reference body composition measurements.

[0030] In some aspects, the techniques described herein relate to a method, wherein the physiological consistency relationships require that predicted compartment masses sum to a value within a predefined tolerance of the subject's measured body weight.

[0031] In some aspects, the techniques described herein relate to a method, wherein the interdependent output variables arise from a constraint-enforcing architecture wherein: during training, the composite loss function penalizes constraint violations causing the model to learn parameter values that produce constraint-satisfying outputs; and during inference, the model produces percentage-based outputs for each compartment that are constrained to sum to one hundred percent before multiplication by measured body weight to yield absolute mass values.

[0032] In some aspects, the techniques described herein relate to a method, wherein the body composition compartments further include one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

[0033] In some aspects, the techniques described herein relate to a method, wherein the method is performed without computing any intermediate anthropometric measurements from the digital images.

[0034] In some aspects, the techniques described herein relate to a method, wherein the method is performed through direct analysis of two-dimensional images without intermediate three-dimensional body reconstruction, thereby eliminating depth-estimation error sources and reducing computational complexity compared to volumetric reconstruction approaches while enabling multi-compartment differentiation that densitometric methods cannot achieve.

[0035] In some aspects, the techniques described herein relate to a method, wherein inference is performed without computing body volume or body density as intermediate values.

[0036] In some aspects, the techniques described herein relate to a method, wherein the trained computational model enforces physiological constraints during model training through the composite loss function rather than through post-processing adjustment or temporal filtering of model outputs.

[0037] In some aspects, the techniques described herein relate to a computer-implemented method for tracking body composition changes in a subject over time, the method performed by a computing system including at least one processor and non-transitory memory, the method including: receiving a first set of digital images of the subject captured at a first time point; determining a first set of body composition metrics from the first set of digital images using a trained computational model; receiving a second set of digital images of the subject captured at a second time point temporally separated from the first time point by a time interval ranging from one day to three hundred sixty-five days; determining a second set of body composition metrics from the second set of digital images using the trained computational model; computing changes between the first set and the second set of body composition metrics, wherein the computed changes include coordinated changes across at least three physiologically distinct body composition compartments whose values remain constrained by physiological consistency relationships; computing a confidence metric for the determined body composition metrics at the second time point, wherein the confidence metric decreases as the time interval increases; and generating a recommendation for obtaining a reference calibration measurement when the confidence metric falls below a predetermined threshold; wherein the confidence metric indicates quantified prediction reliability.

[0038] In some aspects, the techniques described herein relate to a method, wherein the prediction reliability is a function of time elapsed since reference calibration.

[0039] In some aspects, the techniques described herein relate to a method, wherein computing the confidence metric includes analyzing historical validation data showing prediction accuracy degradation as a function of time elapsed since reference measurement.

[0040] In some aspects, the techniques described herein relate to a method, wherein the confidence metric includes a mean absolute error estimate that increases with the time interval.

[0041] In some aspects, the techniques described herein relate to a method, wherein the recommendation may be generated when the time interval exceeds ninety days.

[0042] In some aspects, the techniques described herein relate to a method, further including generating a health risk indicator, wherein the health risk indicator includes a risk assessment for at least one condition selected from the group consisting of: osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolic syndrome, and non-alcoholic fatty liver disease.

[0043] In some aspects, the techniques described herein relate to a method, further including generating an alert when a computed change in at least one body composition metric exceeds a predetermined threshold.

[0044] In some aspects, the techniques described herein relate to a method, wherein the trained computational model was trained on longitudinal data including paired images and medical device measurements from subjects at multiple time points, enabling the model to predict physiologically realistic body composition trajectories.

[0045] In some aspects, the techniques described herein relate to a method, wherein computing changes further includes differentiating between transient changes attributable to hydration status and structural changes attributable to fat mass or lean mass modification.

[0046] In some aspects, the techniques described herein relate to a method, wherein the trained computational model performs inference without computing total body volume, body density, anatomical keypoints, or scaled external body feature dimensions as intermediate values.

[0047] In some aspects, the techniques described herein relate to a system for body composition analysis including: a computer-readable memory storing computer-executable instructions; and one or more processors that execute the computer-executable instructions to at least: receive a set of digital images including a plurality of views of a subject in a standardized anatomical position; isolate the subject from the image background to generate a set of modified images; apply the modified images as input to a trained machine learning model to cause themachine learning model to generate a set of predictions for at least three physiologically distinct compartments based on jointly inferring predicted values for a plurality of body composition metrics, wherein the predicted values are generated by the trained machine learning model based on pixel-level analysis of the modified images, wherein the set of predictions are interdependent, and wherein at least a first prediction of the set of predictions constrains at least a second prediction of the set of predictions; and generate a health assessment based on the set of predictions.

[0048] In some aspects, the techniques described herein relate to a system, further including a mobile computing device including: a second computer-readable memory storing second computer-executable instructions; and a second one or more processors that execute the second computer-executable instructions to at least: present, via a graphical user interface, instructions guiding a user to capture the set of digital images; receive the health assessment; and present, via the graphical user interface, the health assessment.

[0049] In some aspects, the techniques described herein relate to a system, wherein the computer-executable instructions, when executed by the one or more processors, further cause the one or more processors to: generate instructions to cause display of a graphical user interface on a mobile device, wherein the graphical user interface includes instructions for a user to generate the set of images; and transmit the instructions to the mobile device.

[0050] In some aspects, the techniques described herein relate to a system, wherein the mobile computing device provides real-time feedback to guide the user in positioning, framing, and environmental conditions during image capture.

[0051] In some aspects, the techniques described herein relate to a system, wherein the at least three physiologically distinct compartments include fat mass, lean mass, and bone mineral content.

[0052] In some aspects, the techniques described herein relate to a system, wherein the body composition metrics further include one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

[0053] In some aspects, the techniques described herein relate to a system, wherein the health assessment includes diagnostic or medical pre-screening for at least one condition selected from the group consisting of: osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolic syndrome, non-alcoholic fatty liver disease, polycystic ovary syndrome, hypothyroidism, and chronic kidney disease.

[0054] In some aspects, the techniques described herein relate to a system, wherein the computer-executable instructions further cause the one or more processors to track the predicted body composition metrics over time and generate trend analyses and alerts based on detected changes.

[0055] In some aspects, the techniques described herein relate to a system, wherein the trained machine learning model was trained on paired data including digital images captured within a temporal window corresponding to body composition reference measurements from at least one medical imaging device including Dual-Energy X-ray Absorptiometry (DEXA).

[0056] In some aspects, the techniques described herein relate to a system, wherein the body composition metrics further include regional measurements for body segments including at least two of: right arm, left arm, right leg, left leg, and trunk.

[0057] In some aspects, the techniques described herein relate to a system, wherein isolating the subject from the image background includes removing image backgrounds while preserving subject visual information through image segmentation techniques.

[0058] In some aspects, the techniques described herein relate to a system, wherein the one or more processors further resize isolated subject images to standardized dimensions compatible with the trained machine learning model architecture.

[0059] In some aspects, the techniques described herein relate to a method for generating physiologically-constrained body composition predictions over time, the method including: receiving a first set of digital images of a subject captured at a first time point; determining a first set of body composition metrics from the first set of digital images using a trained computational model, wherein the first set of body composition metrics includes at least fat mass percentage and lean mass percentage; receiving a second set of digital images of the subject captured at a second time point temporally separated from the first time point by a time interval; computing a physiologically feasible range of body composition values at the second time point based on the body composition metrics determined at the first time point, the time interval, and physiological constraints on maximum rate of body composition change derived from metabolic limits; determining a second set of body composition metrics from the second set of digital images using the trained computational model; comparing the second set of body composition metrics against the physiologically feasible range; and when the second set of body composition metrics falls outside the physiologically feasible range, adjusting at least one of thepredicted body composition values to fall within the physiologically feasible range, or a confidence metric associated with the prediction to reflect increased uncertainty.

[0060] In some aspects, the techniques described herein relate to a method, wherein the physiologically feasible range expands as the time interval increases, forming a bounded region of physiologically possible body composition changes.

[0061] In some aspects, the techniques described herein relate to a method, wherein adjusting the confidence metric includes increasing an uncertainty bound proportional to the degree by which the predicted values exceed the physiologically feasible range.

[0062] In some aspects, the techniques described herein relate to a method, further including generating a recommendation for obtaining a reference calibration measurement when the confidence metric falls below a predetermined threshold.

[0063] In some aspects, the techniques described herein relate to a method, wherein the confidence metric is computed based on elapsed time since the subject's most recent reference imaging device measurement.

[0064] In some aspects, the techniques described herein relate to a method, wherein the physiological constraints on maximum rate of fat mass change are derived from published metabolic equations defining minimum achievable body fat percentage.

[0065] In some aspects, the techniques described herein relate to a non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to: receive a set of digital images including a plurality of views of a subject in a standardized anatomical position; preprocess the digital images to isolate the subject from image backgrounds; apply a trained machine learning model to the preprocessed images to generate predicted values for at least three physiologically distinct body composition compartments, wherein the trained machine learning model was trained on paired data including digital images captured within a temporal window of body composition reference measurements from at least one medical imaging device, and wherein training used a loss function enforcing physiological mass-balance constraints; and output the predicted values, wherein the predicted values are interdependent such that predicted compartment masses sum to a value within a predefined tolerance of the subject's measured body weight.BRIEF DESCRIPTION OF THE DRAWINGS

[0066] FIG. 1 illustrates a system architecture overview showing the relationship between the mobile device, cloud infrastructure, and analysis engine components.

[0067] FIG. 2 illustrates a data flow diagram showing the processing pipeline from image capture through body composition output generation.

[0068] FIG. 3 illustrates the standardized four-view image capture protocol showing anterior, posterior, left lateral, and right lateral views with anatomical positioning requirements.

[0069] FIG. 4 illustrates the training data collection methodology showing temporal alignment between smartphone image capture and medical device reference measurements.

[0070] FIG. 5 illustrates prediction confidence degradation as a function of time elapsed since reference calibration measurement, showing the relationship between temporal separation and prediction reliability.

[0071] FIG. 6 illustrates a computing device architecture showing processor, memory, computer readable medium drive, network interface, input / output device interface, and associated components.

[0072] FIG. 7 illustrates the composite loss function architecture used during model training, showing the prediction error term, physiological constraint penalty term, weighted combination, and backpropagation to model parameters.DEFINITIONS

[0073] The following definitions apply throughout this specification:

[0074] "Body composition compartments" or "physiologically distinct body composition compartments" refers to categories of body tissue or substance that may be separately quantified in body composition analysis. Examples include fat mass, lean mass, bone mineral content, total body water, visceral adipose tissue, and subcutaneous adipose tissue.

[0075] "Direct inference" refers to a computational approach wherein body composition metrics may be predicted directly from image pixel data or learned image feature representations, without computing intermediate geometric, volumetric, or anthropometric quantities as explicit steps.

[0076] "Joint inference" refers to a prediction approach wherein multiple body composition metrics may be determined through a unified computational process that considers their interdependencies, rather than through separate independent predictions for each metric.

[0077] "Loss function" refers to the objective function minimized during machine learning model training. A loss function that penalizes violations of physiological mass-balance constraints may include terms that increase loss when predicted compartment values fail to satisfy mass-balance relationships.

[0078] "Interdependent output constraints " refers to a prediction architecture wherein the values generated for multiple output variables may be jointly determined rather than independently computed, such that changing the predicted value for one compartment necessarily affects the predicted values for other compartments.

[0079] "Mean absolute error” or "MAE" refers to the average of absolute differences between predicted values and reference values across a validation dataset.

[0080] "Medical imaging device" refers to diagnostic equipment designed and validated for measuring body composition. Examples include: Dual-Energy X-ray Absorptiometry (DEXA) systems, Bioelectrical Impedance Analysis (BIA) devices, Air Displacement Plethysmography (ADP) systems, Magnetic Resonance Imaging (MRI) systems, and Computed Tomography (CT) scanners, or any other device capable of generating an image representing at least a portion of the body

[0081] "Physiological consistency relationships” refers to constraints that may ensure predicted body composition values are compatible with known biological and physical principles, including compartment masses summing to total body weight and individual compartment values falling within biologically feasible ranges.

[0082] "Physiological mass-balance constraints" refers to constrained relationships that may require predicted body composition compartment values to satisfy fundamental physical and biological principles, including requirements that compartment masses sum to total body weight within a predefined tolerance.

[0083] "Pixel-level analysis of two-dimensional images” refers to computational processing that operates on the pixel values of standard two-dimensional photographs, including convolutional neural network processing or other computer vision techniques.

[0084] "Surface-geometry-based anthropometric measurements" refers to body dimension measurements derived from the external surface geometry of the body, such as waist circumference, hip circumference, and limb lengths.

[0085] "Temporal window " or "temporal alignment" refers to the maximum allowable time separation between capture of smartphone images and acquisition of corresponding reference measurements. Suitable temporal windows may include, for example, thirty seconds, five minutes, ten minutes, thirty minutes, one hour, six hours, twelve hours, one day, or one week, or other durations appropriate for the specific measurement context.

[0086] "Three-dimensional body reconstruction" refers to computational processes that generate explicit three-dimensional geometric representations of a subject's body, including point clouds, surface meshes, or parametric body models.

[0087] "Physiologically feasible range " refers to the bounded set of body composition values achievable within a given time interval based on metabolic constraints. For example, under high-intensity intervention conditions, body fat percentage may change by approximately 0.5% to 1.5% per week; under moderate intervention conditions, approximately 0.2% to 0.5% per week; under low-intensity or maintenance conditions, less than 0.2% per week.DETAILED DESCRIPTION1. Limitations of Existing Body Composition Analysis Systems

[0088] Several medical device-based body composition assessment methods exist, however each has limitations that reduce accessibility for routine health monitoring and preventive care applications.1.1 Medical Imaging Device Limitations

[0089] Dual-Energy X-ray Absorptiometry (DEXA) provides bone density, lean mass, and fat mass measurements with high accuracy. However, DEXA systems require dedicated facilities with specialized power systems and controlled environments, involve low-level ionizing radiation exposure that limits measurement frequency, and necessitate trained technician operation. These requirements restrict DEXA availability to clinical settings and make routine longitudinal monitoring impractical for most individuals.

[0090] Bioelectrical Impedance Analysis (BIA) measures total body water and estimates lean mass through electrical conductance properties. While more accessible than DEXA,BTA accuracy depends heavily on hydration status, recent food intake, and ambient temperature. BIA provides limited differentiation between tissue compartments and cannot directly measure bone mineral content or distinguish between visceral and subcutaneous adipose tissue.

[0091] Air Displacement Plethysmography (ADP) determines body density and derives body fat percentage through volumetric displacement measurement. ADP requires specialized chamber equipment with precise environmental controls, limiting deployment to dedicated facilities. The densitometric approach provides only two-compartment analysis (fat mass versus fat-free mass) without finer tissue differentiation.

[0092] Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) provide detailed internal tissue structure visualization. These modalities require substantial capital equipment investment, dedicated facilities, and trained operators. CT involves significant radiation exposure inappropriate for routine monitoring. Both technologies have contraindications for certain populations and require extended measurement protocols incompatible with frequent longitudinal tracking.1.2 Mobile Application Limitations

[0093] Existing smartphone-based body composition applications have attempted to address accessibility limitations of medical imaging devices. However, current implementations rely on approaches that fundamentally limit their accuracy and clinical utility.

[0094] Some existing approaches use anatomical key point detection to identify joint locations and body landmarks, then compute scaled external body dimensions such as waist circumference or limb lengths. These surface-geometry-based anthropometric measurements are then used in regression equations to estimate body composition. This approach inherits the well-documented limitations of anthropometric prediction equations, which have limited accuracy across diverse populations and cannot distinguish between internal tissue compartments.

[0095] Other existing approaches attempt three-dimensional body reconstruction from multiple images, computing total body volume and deriving body density for densitometric body composition estimation. This volumetric approach requires accurate depth estimation and surface reconstruction, introducing substantial error sources. The densitometric derivation provides only two-compartment analysis and cannot differentiate between specific tissue types such as visceral versus subcutaneous adipose tissue or quantify bone mineral content.

[0096] Existing mobile applications also typically lack validation against medical imaging standards, operate without physiological consistency constraints that would ensure predicted values are biologically plausible, and do not leverage longitudinal medical device data for training.

[0097] The systems and methods of the present disclosure address these technical limitations by providing at least: (1) a training methodology that enforces physiological constraints through loss function design rather than post-processing, causing the trained model to inherently produce constraint-satisfying outputs; (2) a direct inference architecture that eliminates intermediate reconstruction steps and their associated error propagation; (3) a temporal window requirement for training data that ensures reference measurements accurately reflect subject state at image capture time; and (4) controlled environmental variability during training that produces models robust to real-world deployment conditions. These technical features, alone or in combination, provide a technical improvement over existing approaches by achieving accurate, comprehensive, multi-compartment body composition inference from smartphone images at accuracy levels that may be validated against medical imaging devices.2. Unmet Needs in Body Composition Monitoring

[0098] The limitations of existing systems create significant unmet needs for patients, clinicians, and researchers requiring accessible body composition monitoring.2.1 Clinical Need for Accessible Longitudinal Monitoring

[0099] Numerous chronic conditions would benefit from regular body composition monitoring, yet current medical imaging approaches cannot practically support the monitoring frequency that optimal clinical management requires. Type 2 diabetes management, for instance, benefits from tracking visceral adipose tissue changes in response to lifestyle modifications and pharmaceutical interventions, but quarterly or monthly DEXA scans are neither practical nor cost-effective. Similarly, osteoporosis treatment monitoring ideally requires bone density assessment to evaluate pharmaceutical intervention effectiveness over time, yet radiation exposure concerns limit DEXA frequency. Sarcopenia screening in aging populations benefits from skeletal muscle mass quantification at intervals that enable early intervention, while cancer cachexia monitoring requires lean mass tracking frequent enough to guide timely nutritional intervention.

[0100] The emergence of GLP- 1 receptor agonist therapies for weight management has created particular demand for accessible body composition monitoring. These medicationsproduce significant weight loss, but clinical outcomes depend critically on the composition of that weight loss - specifically, the ratio of fat mass reduction to lean mass preservation. Patients and clinicians require frequent body composition assessment to optimize dosing, adjust nutritional protocols, and ensure metabolically healthy weight loss trajectories. Current medical imaging infrastructure cannot support the monitoring frequency these therapies demand.

[0101] A system enabling accessible, radiation-free, multi-compartment body composition analysis would address substantial unmet clinical need across these and other applications.2.2 Need for Multi-Compartment Analysis

[0102] Two-compartment body composition models distinguishing only fat mass from fat-free mass provide insufficient granularity for many clinical applications. The ability to separately quantify bone mineral content, skeletal muscle mass, visceral adipose tissue, subcutaneous adipose tissue, and total body water enables substantially more precise health risk assessment and treatment guidance than aggregate two-compartment measures permit.

[0103] Visceral adipose tissue accumulation correlates with metabolic syndrome and cardiovascular disease risk independent of total body fat percentage, making VAT-specific quantification clinically valuable for risk stratification. Bone mineral content trajectories indicate osteoporosis risk and enable preventive intervention before fracture events occur. Total body water measurement enables differentiation between acute hydration fluctuations and structural body composition changes, improving the clinical interpretability of longitudinal measurements. Regional muscle mass distribution affects mobility, fall risk, and functional capacity in aging populations, with clinical implications distinct from total lean mass measures.

[0104] Existing accessible solutions using densitometric or anthropometric approaches cannot provide this multi-compartment differentiation. A system capable of inferring multiple physiologically distinct compartments from accessible imaging would address this fundamental limitation.2.3 Need for Physiologically Consistent Predictions

[0105] Existing machine learning approaches to body composition prediction typically treat each output metric independently, without constraints ensuring the predicted values are physiologically plausible. Independent prediction can produce results where compartment massesdo not sum to total body weight, where changes between successive measurements violate conservation of mass, or where compartment ratios fall outside biologically realistic ranges.

[0106] Clinically useful body composition analysis requires outputs that satisfy physiological consistency relationships. A system producing interdependent predictions where compartment values mutually constrain each other would provide more reliable results suitable for clinical decision support.3. System Overview

[0107] The present disclosure provides systems and methods for multi-compartment body composition analysis that address the limitations of existing approaches and fulfill the identified unmet needs. The system combines accessible smartphone-based image capture with machine learning models trained on paired data from medical imaging devices, using a training methodology that enforces physiological consistency constraints.

[0108] The system achieves multi-compartment body composition inference directly from digital images without requiring three-dimensional body reconstruction, body volume computation, body density calculation, anatomical keypoint detection, scaled external body dimension estimation, or surface-geometry-based anthropometric measurements. This direct inference approach, combined with physiological constraint enforcement during model training, enables accurate multi-compartment prediction from accessible imaging.

[0109] Referring to FIG. 1, the system comprises a mobile application (102) executing on a mobile computing device (101), cloud infrastructure (103) including a preprocessing module (104) and trained computational model (105), and a reporting module (106). The mobile application guides users through standardized image capture. Captured images are transmitted to the cloud infrastructure for preprocessing and analysis. The trained computational model applies trained models to generate body composition predictions, which are returned via the reporting module for display and longitudinal tracking.4. Image Capture Protocol

[0110] Referring to FIG. 3, the system implements a standardized four-view image capture protocol designed to provide sufficient visual information for accurate body composition inference while remaining practical for routine self-administration.4.1 Anatomical Positioning

[0111] The subject assumes a standardized anatomical position with feet positioned slightly wider than shoulder-width apart and arms held between thirty and sixty degrees away from the body. This positioning ensures consistent visualization of body contours and minimizes occlusion of relevant anatomical regions across the four captured views.4.2 Four- View Capture

[0112] The capture protocol comprises four views: an anterior view (front-facing), a posterior view (back-facing), a left lateral view (left side profile), and a right lateral view (right side profile). The mobile application provides real-time guidance to assist users in achieving proper positioning, framing, and capture timing for each view.

[0113] The four-view protocol provides complementary visual information enabling inference of three-dimensional body composition distribution from two-dimensional images. The anterior and posterior views capture torso width and limb positioning. The lateral views capture anterior-posterior depth, abdominal protrusion indicative of visceral adipose tissue distribution, and spinal curvature.

[0114] Referring to FIG. 3, the capture protocol comprises four views: an anterior view (301), a posterior view (302), a right lateral view (303), and a left lateral view (304). The subject assumes a standardized anatomical position with arms angled away from the body and feet in a shoulder-width stance.

[0115] Further, additional physical measurements and features may be determined from the captured images. The computational model may learn to extract features from combinations of body regions including torso geometry and facial characteristics to inform body composition predictions.4.3 Controlled Environmental Variability

[0116] The training data collection protocol intentionally introduces controlled environmental variability across captured image sets. This variability comprises variation in at least two of: image background, ambient lighting conditions, device orientation angle, shadow characteristics, and subject clothing.

[0117] In one implementation, training data is collected using a protocol of five cycled backgrounds, each with different lighting conditions, shadow patterns, device angles, and subject-to-device distances. This intentional variability during training enables the resulting model togeneralize across the diverse real-world capture conditions encountered during deployment, rather than overfitting to specific environmental configurations.

[0118] The training process may further incorporate data augmentation techniques including, but not limited to, random rotation, scaling, color jittering, horizontal flipping, brightness adjustment, contrast modification, and additive noise injection. These augmentation techniques, in combination with the controlled environmental variability during capture, enhance the model's ability to generalize across deployment conditions.5. Training Methodology

[0119] The system employs a training methodology that pairs smartphone images with temporally-constrained body composition measurements from medical imaging devices, using a loss function that enforces physiological mass-balance constraints.5.1 Paired Training Data Collection

[0120] Referring to FIG. 4, training data comprises paired datasets where smartphone image sets are captured within a defined temporal window of body composition reference measurements from at least one medical imaging device. In various implementations, the temporal window may be thirty seconds, five minutes, ten minutes, thirty minutes, one hour, or another duration selected based on the precision requirements of the application

[0121] Reference measurements may be obtained from one or more devices selected from Dual-Energy X-ray Absorptiometry (DEXA), Bioelectrical Impedance Analysis (BIA), Air Displacement Plethysmography (ADP), Magnetic Resonance Imaging (MRI), and Computed Tomography (CT). Among these modalities, DEXA may provide suitable reference data because it directly and simultaneously measures bone mineral content, lean mass, and fat mass with established clinical accuracy, enabling training of multi-compartment models without requiring data fusion across multiple reference devices.

[0122] Each training data point associates the captured image set with the corresponding body composition reference measurements and demographic data. The subject’s measured body weight may be provided as a direct input to the inference system for constraint enforcement. Additional demographic data comprising at least two of: age, biological sex, height, and ethnicity may be associated with training data for population stratification and validation purposes.

[0123] Referring to FIG. 4, training data comprises paired datasets where medical device measurements (401) and smartphone image capture (402) occur within a defined temporal window. In one implementation, the temporal window may be ten minutes or less. The temporal alignment ensures reference measurements accurately reflect the subject's body composition at the time of image capture. The medical device measurement and smartphone images are associated to form paired training data (403).5.2 Longitudinal Training Data

[0124] The training dataset includes longitudinal paired data comprising multiple image sets and corresponding reference measurements for individual training subjects captured at different time points. In one implementation, longitudinal data points are separated by intervals of approximately four to eight weeks, sufficient duration for physiologically meaningful body composition changes to occur.

[0125] Longitudinal training data enables the model to learn physiologically realistic body composition trajectories, distinguishing between plausible body composition changes and measurement artifacts. This longitudinal constraint improves prediction stability and supports the system's longitudinal tracking capabilities.5.3 Image Preprocessing

[0126] Prior to model training and inference, captured images undergo preprocessing to isolate the subject from the image background. This preprocessing step removes environmental variability in backgrounds while preserving the subject visual information relevant to body composition inference.

[0127] In one implementation, background removal may be performed using established image segmentation techniques to separate the subject from the capture environment. The isolated subject images are then resized to standardized dimensions compatible with the computational model architecture. These preprocessing steps prepare consistent input data for the trained model while remaining agnostic to specific segmentation implementations.5.4 Physiological Mass-Balance Constraint Training

[0128] The computational model may be trained using a composite loss function that penalizes violations of physiological mass-balance constraints. These constraints require that inferred body composition compartment values satisfy interdependency relationships reflecting physiological reality.

[0129] Referring to FIG. 7, the composite loss function architecture comprises model outputs (701) representing predicted compartment percentages, and reference values (702) obtained from medical imaging devices during training data collection. A prediction error term (703) measures the deviation between the model outputs and reference values, typically computed as mean squared error or mean absolute error across the training batch. A sum constraint check (704) verifies whether predicted compartment percentages sum to one hundred percent. A constraint penalty term (705) increases the loss value when predicted percentages violate the sum constraint, with penalty magnitude proportional to the degree of violation. A weighted combination (706) combines the prediction error term and constraint penalty term to produce a composite loss value (707). During training, backpropagation (708) computes gradients of the composite loss with respect to model parameters, enabling iterative parameter adjustment that minimizes both prediction error and constraint violations.

[0130] In one implementation, the computational model predicts body composition compartments as percentages of total body mass. The user provides measured body weight as an input, and the predicted percentages are constrained to sum to one hundred percent of body mass. This hard constraint ensures that predicted compartment values, when converted to absolute masses by multiplication with the provided body weight, necessarily sum to total body weight. The constraint applies both during model training and at inference time.

[0131] This percentage-based prediction with hard constraint enforcement produces interdependent outputs such that the predicted value for any one compartment constrains the predicted values for remaining compartments. Unlike unconstrained machine learning approaches where each output is predicted independently, the constraint-trained model guarantees that compartment relationships satisfy mass-balance requirements. If the model predicts a higher percentage for one compartment, the remaining compartments may correspondingly decrease to maintain the sum constraint.5.5 Direct Inference Without Intermediate Reconstruction

[0132] The computational model may be trained to infer body composition metrics directly from digital images without requiring intermediate computational steps that characterize prior approaches. Specifically, the model performs inference without: (i) reconstructing a three-dimensional body model, (ii) computing total body volume or body density as intermediate values,(iii) generating anatomical keypoints, (iv) computing scaled external body feature dimensions, or (v) deriving surface-geometry-based anthropometric measurements.

[0133] This direct inference approach distinguishes the present system from prior art approaches relying on volumetric densitometry, keypoint-based anthropometry, or surface geometry analysis. By learning direct mappings from image features to body composition metrics during training on paired medical device data, the model avoids error propagation through intermediate reconstruction steps and can capture relationships not expressible through anthropometric equations.6. Body Composition Inference6.1 Multi-Compartment Output

[0134] The trained computational model infers body composition metrics comprising at least three physiologically distinct body composition compartments. In a base implementation, these compartments comprise fat mass, lean mass, and bone mineral content. In expanded implementations, the body composition compartments further comprise one or more of: total body water, visceral adipose tissue volume, bone mineral density, and subcutaneous adipose tissue volume.

[0135] This multi-compartment output provides substantially greater clinical utility than two-compartment models that distinguish only fat mass from fat-free mass. Visceral adipose tissue quantification enables cardiovascular and metabolic risk stratification independent of total adiposity, as VAT accumulation carries distinct health implications from subcutaneous fat distribution. Total body water measurement enables clinicians and users to distinguish acute hydration fluctuations from structural tissue changes, improving interpretability of longitudinal measurements particularly in contexts such as athletic training or medication effects. Bone mineral content and density assessment supports osteoporosis risk screening and enables monitoring of bone health trajectories that may indicate need for preventive intervention.6.2 Interdependent Output Variables

[0136] The trained model produces interdependent output variables for the plurality of body composition compartments such that the predicted value for any one compartment constrains the predicted values for remaining compartments. This interdependency arises from the physiological mass-balance constraint enforcement described above.

[0137] In one implementation, the computational model processes the input images to extract feature representations relevant to different body composition compartments. These feature representations inform joint prediction of compartment percentages, where the percentage outputs are constrained to sum to one hundred percent. The joint inference architecture ensures that predicted compartment values satisfy physiological consistency relationships, as increasing the predicted percentage for one compartment necessarily decreases the available percentage allocation for remaining compartments.

[0138] The specific neural network architecture for feature extraction and output generation may vary across implementations. The interdependency of outputs arises from the constraint enforcement rather than from any particular architectural choice, enabling flexibility in model design while maintaining the physiological consistency guarantees that distinguish the system from unconstrained prediction approaches.6.3 Regional Body Composition

[0139] In some implementations, the system outputs regional body composition measurements for body segments selected from the group consisting of: right arm, left arm, right leg, left leg, and trunk. Regional measurements enable assessment of body composition symmetry, detection of localized muscle wasting indicative of specific conditions, and tracking of regional response to targeted interventions.6.4 Accuracy Validation

[0140] The system achieves body composition prediction accuracy validated against medical imaging device measurements. In one implementation, predicted values for fat mass demonstrate a mean absolute error of less than five percent when evaluated against measurements from Dual-Energy X-ray Absorptiometry (DEXA). Bone mineral density predictions demonstrate statistically significant predictive signal, enabling population-level screening applications.

[0141] Validation studies demonstrate systematic performance characteristics across the training dataset progression. In datasets comprising approximately 750 paired samples, body fat percentage prediction achieves mean absolute error in the range of 3.0% to 3.3% when validated against DEXA measurements, with coefficient of determination (R2) values ranging from 0.69 to 0.85. Performance demonstrates temporal dependency, with prediction accuracy highest for measurements obtained within 90 days of reference calibration (MAE approximately 2.5% to3.0%) and degrading systematically as time interval increases (MAE approximately 3.2% to 3.5% at 12 months post-reference).

[0142] Bone mineral density predictions demonstrate preliminary correlation with DEXA measurements (R2approximately 0.35-0.40 in datasets of 400-750 paired samples), sufficient for population-level screening applications and risk stratification. This correlation demonstrates systematic improvement trajectory as training dataset size increases, indicating that continued data collection will enhance bone density prediction accuracy toward levels suitable for clinical diagnostic applications.

[0143] Prediction accuracy may vary across demographic subgroups depending on training data composition. In implementations with initial training datasets, body fat percentage prediction for male subjects may achieve coefficient of determination (R2) values of approximately 0.75 to 0.80, while female subjects may achieve R2values of approximately 0.35 to 0.50 with smaller training populations. This differential may be addressed through continued data collection emphasizing demographic balance, implementation of sex-specific model variants that apply different trained parameters based on subject demographic characteristics, and application of transfer learning techniques. The system may automatically select appropriate model variants based on user- provided demographic inputs.7. Longitudinal Body Composition Tracking

[0144] The system supports longitudinal body composition tracking by processing image sets captured at multiple time points and computing changes between successive measurements.7.1 Change Computation

[0145] Longitudinal tracking comprises receiving a first set of digital images captured at a first time point, determining a first set of body composition metrics, receiving a second set of digital images captured at a second time point temporally separated from the first, determining a second set of body composition metrics, and computing changes between the first and second sets.

[0146] The computed changes comprise coordinated changes across at least three physiologically distinct body composition compartments whose values are interdependent and remain constrained by physiological consistency relationships over time. This coordinated change computation ensures that detected changes reflect physiologically plausible body composition trajectories.7.2 Trajectory Analysis

[0147] The system analyzes body composition trajectories across all measured compartments to identify clinically meaningful patterns and distinguish between different types of body composition changes. Fat mass trajectories may indicate response to dietary interventions, exercise regimens, or pharmacological treatments. Lean mass trajectories reveal muscle preservation or loss patterns relevant to sarcopenia risk, athletic training adaptation, or medication side effects. Bone mineral content trajectories over extended time periods may indicate osteoporosis progression or treatment response. Total body water changes help differentiate acute hydration fluctuations from structural tissue modifications, improving the interpretability of shortterm measurements.

[0148] The system may identify trajectory patterns characteristic of specific clinical scenarios. Rapid lean mass decline concurrent with stable or increasing fat mass may indicate sarcopenic obesity risk. Coordinated reduction in fat mass with preservation of lean mass characterizes metabolically favorable weight loss. Divergent trajectories between measured and expected compartment changes based on caloric balance may indicate metabolic adaptation or measurement timing artifacts.7.3 Health Risk Indicators

[0149] The system generates health risk indicators based on analysis of body composition values and trajectories. Risk assessment incorporates both absolute compartment values relative to population norms stratified by age, sex, and other demographic factors, and trajectory analysis identifying concerning rates or patterns of change.

[0150] For osteoporosis and osteopenia risk, the system evaluates bone mineral content and density values against age and sex-adjusted reference ranges, and monitors trajectories for accelerated bone loss patterns. For sarcopenia risk, the system assesses skeletal muscle mass relative to established thresholds and identifies trajectories suggesting accelerated age-related muscle loss. For metabolic syndrome and cardiovascular risk, the system evaluates visceral adipose tissue accumulation patterns and VAT-to-subcutaneous fat ratios. For cachexia screening in relevant populations, the system monitors for concurrent lean mass and fat mass depletion patterns.

[0151] Health risk indicators may comprise risk assessments for conditions selected from the group consisting of: osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolicsyndrome, non-alcoholic fatty liver disease, polycystic ovary syndrome, hypothyroidism, and chronic kidney disease.

[0152] The system may generate alerts when a computed change in at least one body composition metric exceeds a predetermined threshold, or when trajectory analysis identifies patterns associated with elevated health risk. Alerts may prompt user attention, recommend lifestyle modifications, or suggest clinical consultation depending on the nature and severity of the identified pattern.7.4 Prediction Confidence, Recalibration, and Temporal Uncertainty

[0153] The system computes confidence metrics for body composition predictions based on temporal separation from most recent reference calibration measurements. Historical validation data demonstrates that prediction accuracy degrades systematically as time interval increases between reference measurement and smartphone-based prediction.

[0154] In one implementation, the system tracks time elapsed since a user's most recent reference measurement (if any). Confidence scores are computed based on validation curves showing mean absolute error increase as a function of time elapsed. For users without recent reference measurements, confidence scores reflect population-level accuracy statistics stratified by demographic factors.

[0155] The system may generate recommendations for obtaining reference calibration measurements when confidence scores fall below predetermined thresholds. In one implementation, recalibration may be recommended when time since the last reference exceeds 180 days, ensuring that longitudinal tracking maintains clinical accuracy standards. Upon receiving new reference data, the system may adjust subsequent predictions to incorporate the updated calibration point, reducing prediction uncertainty for future measurements.

[0156] This confidence-aware prediction approach addresses a fundamental limitation of existing body composition monitoring solutions, which typically provide point estimates without quantifying reliability or indicating when measurements could be repeated.

[0157] Referring to FIG. 5, the system computes a confidence metric for body composition predictions that accounts for temporal separation from reference calibration measurements. A reference measurement (501) represents the point of highest prediction confidence, obtained when the subject has a recent medical device measurement. The confidence degradation curve (502) shows that prediction confidence decreases as time elapsed since thereference measurement increases. A recalibration threshold (503) indicates the confidence level below which the system generates a recommendation for obtaining a new reference calibration measurement. The recalibration recommendation zone (504) indicates the region where prediction confidence has fallen below the threshold and recalibration is recommended.

[0158] In implementations incorporating physiologically-constrained predictions, the system defines a physiologically feasible range of body composition values that expands as time elapsed increases, reflecting the accumulated uncertainty in possible body composition trajectories. This expanding range (506) forms a bounded region of physiologically possible body composition changes based on metabolic constraints on maximum rates of fat mass loss and lean mass modification.8. System Architecture

[0159] Referring to FIG. 1, the system architecture comprises a mobile application, preprocessing module, server system, and reporting module.

[0160] Referring to FIG. 2, the inference pipeline comprises image capture (201), preprocessing for subject isolation (202), application of the trained computational model (203), and body composition output generation (204). The subject's measured body weight (205) is provided as a separate input to the trained computational model to enable physiological massbalance constraint enforcement, wherein predicted compartment percentages are multiplied by the measured body weight to produce absolute mass values that sum to total body weight.8.1 Mobile Application

[0161] The mobile application executes on a mobile computing device and may guide a user through capture of a set of digital images comprising a plurality of views of a subject in a standardized anatomical position. The mobile application provides real-time feedback to guide the user in positioning, framing, and environmental conditions during image capture.8.2 Preprocessing Module

[0162] The preprocessing module may isolate the subject from the image background, preparing captured images for analysis by the trained computational model.8.3 Server System

[0163] The server system comprises at least one processor and memory storing a trained computational model. The server system receives preprocessed image data, applies thetrained computational model to perform joint inference, and returns predicted body composition metrics.8.4 Reporting Module

[0164] The reporting module may generate health assessments based on the predicted values for the plurality of body composition metrics. The reporting module may track predicted body composition metrics over time and generate trend analyses and alerts based on detected changes.The reporting module may generate diagnostic screening assessments for conditions including osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolic syndrome, non-alcoholic fatty liver disease, polycystic ovary syndrome, hypothyroidism, and chronic kidney disease.8.5 Computing Device Architecture

[0165] Referring to FIG. 6, the server system (or mobile computing device in edge deployment configurations) comprises a computing device (600) including a computer processor (601), which may be, for example, a central processing unit (CPU), application-specific integrated circuit (ASIC), or field-programmable gate array (FPGA). The computing device further comprises a network interface (602), a computer readable medium drive (603) which may be, for example, a hard disk drive, solid state storage device, CD-ROM, or any other storage medium, an input / output device interface (604), and memory (605). The memory may comprise random access memory (RAM) and stores an operating system (606), the machine learning system (607) comprising the trained computational model, and an information analysis system (608).9. Exemplary Use Cases

[0166] The following non-limiting examples illustrate applications of the disclosed system.9.1 Personal Health and Fitness Monitoring

[0167] An individual uses the smartphone application to capture periodic body composition scans and track progress toward health or fitness goals. The application provides personalized insights based on body composition trajectories and alerts the user to noteworthy changes warranting attention.9.2 GLP-1 Therapy Monitoring

[0168] A patient prescribed GLP-1 receptor agonist medication for weight management uses the system to monitor body composition changes throughout treatment. The system tracks the ratio of fat mass loss to lean mass change, enabling the patient and their healthcare provider to assess whether weight loss may be occurring in a metabolically favorable composition. If trajectory analysis indicates disproportionate lean mass loss, the care team may adjust nutritional protocols, exercise recommendations, or medication dosing to optimize treatment outcomes. The accessible monitoring frequency enabled by the system supports more responsive treatment optimization than periodic clinical DEXA scans would permit.9.3 Clinical Screening

[0169] A healthcare provider deploys the system for sarcopenia screening in elderly patients during routine visits. The system's ability to quantify skeletal muscle mass enables early detection of muscle loss, facilitating timely intervention to reduce frailty and disability risk.9.4 Disease Management

[0170] An endocrinology practice uses the system to monitor patients with type 2 diabetes and metabolic syndrome. Regular body composition assessments guide treatment decisions based on changes in visceral adiposity and lean mass response to interventions.9.5 Research Applications

[0171] Clinical researchers employ the system as a body composition endpoint in pharmaceutical trials. The accessible measurement methodology enables more frequent longitudinal data collection than traditional medical imaging approaches permit.TERMINOLOGY

[0172] All of the methods and tasks described herein may be performed and fully automated by a computer system. The computer system may. in some cases, include multiple distinct computers or computing devices (e.g., physical servers, workstations, storage arrays, cloud computing resources, etc.) that communicate and interoperate over a network to perform the described functions. Each such computing device typically includes a processor (or multiple processors) that executes program instructions or modules stored in a memory or other non-transitory computer-readable storage medium or device (e.g., solid state storage devices, disk drives, etc.). The various functions disclosed herein may be embodied in such program instructions, or may be implemented in application- specific circuitry (e.g., ASICs or FPGAs) ofthe computer system. Where the computer system includes multiple computing devices, these devices may, but need not, be co-located. The results of the disclosed methods and tasks may be persistently stored by transforming physical storage devices, such as solid-state memory chips or magnetic disks, into a different state. In some embodiments, the computer system may be a cloudbased computing system whose processing resources are shared by multiple distinct business entities or other users.

[0173] Depending on the embodiment, certain acts, events, or functions of any of the processes or algorithms described herein can be performed in a different sequence, can be added, merged, or left out altogether (e.g„ not all described operations or events are necessary for the practice of the algorithm). Moreover, in certain embodiments, operations or events can be performed concurrently, e.g., through multi-threaded processing, interrupt processing, or multiple processors or processor cores or on other parallel architectures, rather than sequentially.

[0174] The various illustrative logical blocks, modules, routines, and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, or combinations of electronic hardware and computer software. To clearly illustrate this interchangeability, various illustrative components, blocks, modules, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware, or as software that runs on hardware, depends upon the particular application and design conditions imposed on the overall system. The described functionality can be implemented in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the disclosure.

[0175] Moreover, the various illustrative logical blocks and modules described in connection with the embodiments disclosed herein can be implemented or performed by a machine, such as a processor device, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A processor device can be a microprocessor, but in the alternative, the processor device can be a controller, microcontroller, or state machine, combinations of the same, or the like. A processor device can include electrical circuitry configured to process computer-executable instructions. In another embodiment, a processor device includes an FPGA or other programmable device that performs logic operations withoutprocessing computer-executable instructions. A processor device can also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. Although described herein primarily with respect to digital technology, a processor device may also include primarily analog components. For example, some or all of the algorithms described herein may be implemented in analog circuitry or mixed analog and digital circuitry. A computing environment can include any type of computer system, including, but not limited to, a computer system based on a microprocessor, a mainframe computer, a digital signal processor, a portable computing device, a device controller, or a computational engine within an appliance, to name a few.

[0176] The elements of a method, process, routine, or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor device, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory. EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of a non-transitory computer-readable storage medium. An exemplary storage medium can be coupled to the processor device such that the processor device can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor device. The processor device and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor device and the storage medium can reside as discrete components in a user terminal.

[0177] Conditional language used herein, such as, among others, "can," "could," "might," "may," “e.g.,” and the like, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements and / or steps. Thus, such conditional language is not generally intended to imply that features, elements and / or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without other input or prompting, whether these features, elements and / or steps are included or are to be performed in any particular embodiment. The terms “comprising,” “including,” “having,” and the like are synonymous and are used inclusively, in an open-ended fashion, and do not exclude additional elements, features, acts, operations, and so forth. Also, theterm “or” is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term “or” means one, some, or all of the elements in the list.

[0178] Disjunctive language such as the phrase “at least one of X, Y, Z,” unless specifically stated otherwise, is otherwise understood with the context as used in general to present that an item, term, etc., may be either X, Y, or Z, or any combination thereof (e.g.. X, Y, and / or Z). Thus, such disjunctive language is not generally intended to, and should not, imply that certain embodiments require at least one of X, at least one of Y, or at least one of Z to each be present.

[0179] Unless otherwise explicitly stated, articles such as “a” or “an” should generally be interpreted to include one or more described items. Accordingly, phrases such as “a device configured to” are intended to include one or more recited devices. Such one or more recited devices can also be collectively configured to carry out the stated recitations. For example, “a processor configured to carry out recitations A, B and C” can include a first processor configured to carry out recitation A working in conjunction with a second processor configured to carry out recitations B and C.

[0180] While the above detailed description has shown, described, and pointed out novel features as applied to various embodiments, it can be understood that various omissions, substitutions, and changes in the form and details of the devices or algorithms illustrated can be made without departing from the spirit of the disclosure. As can be recognized, certain embodiments described herein can be embodied within a form that does not provide all of the features and benefits set forth herein, as some features can be used or practiced separately from others. The scope of certain embodiments disclosed herein is indicated by the appended claims rather than by the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

[0181] Reference throughout this specification to “some embodiments” or “an embodiment” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least some embodiments. Thus, appearances of the phrases “in some embodiments” or “in an embodiment” in various places throughout this specification are not necessarily all referring to the same embodiment and may refer to one or more of the same or different embodiments. Furthermore, the particular features, structures or characteristics may becombined in any suitable manner, as would be apparent to one of ordinary skill in the art from this disclosure, in one or more embodiments.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for generating a predictive model for body composition analysis, the method performed by one or more processors executing instructions stored in non-transitory computer- readable memory, the method comprising:for each of a plurality of training subjects, capturing a set of digital images using a mobile computing device, wherein each set comprises a plurality of views of the subject in a standardized anatomical position;introducing controlled environmental variability across the captured image sets, wherein the environmental variability comprises variation in at least two of: image background, ambient lighting conditions, device orientation angle, shadow characteristics, and subject clothing;for each training subject, acquiring body composition reference measurements from at least one medical imaging device selected from the group consisting of: Dual-Energy X- ray Absorptiometry (DEXA), Bioelectrical Impedance Analysis (BIA), Air Displacement Plethysmography (ADP), Magnetic Resonance Imaging (MRI), and Computed Tomography (CT) within a temporal window relative to the image capture, wherein the reference measurements are used for model training and are not required during subsequent inference on new subjects;associating the captured image sets with the corresponding body composition reference measurements and demographic data to form paired training data;preprocessing the captured images to isolate each subject from the image background: andtraining a computational model on the paired training data using backpropagation to iteratively adjust model parameters by minimizing a composite loss function, wherein the composite loss function comprises:a prediction error term measuring deviation between model outputs and reference measurements, and;a physiological constraint term that penalizes violations of mass-balance relationships requiring predicted compartment masses to sum to the subject'smeasured body weight within a predefined tolerance; wherein the trained computational model performs direct inference from two-dimensional image pixel data to body composition outputs without computing intermediate three- dimensional reconstructions, body volume, or surface-geometry-based anthropometric measurements.

2. The method of claim 1, wherein the physiological mass-balance relationships require that inferred compartment values sum to a value within a predefined tolerance of the subject's measured body weight.

3. The method of any of claims 1 to 2, wherein the plurality of views comprises an anterior view, a posterior view, a left lateral view, and a right lateral view.

4. The method of any of claims 1 to 3, wherein the at least one medical imaging device comprises one or more devices capable of generating body composition reference measurements, including devices selected from the group consisting of: Dual-Energy X-ray Absorptiometry (DEXA), Bioelectrical Impedance Analysis (BIA), Air Displacement Plethysmography (ADP), Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and multi-frequency bioimpedance spectroscopy systems.

5. The method of any of claims 1 to 4, wherein the at least three physiologically distinct body composition compartments comprise: fat mass, lean mass, and at least one additional compartment selected from the group consisting of: bone mineral content, total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

6. The method of claim 5, wherein the body composition compartments further comprise one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

7. The method of any of claims 1 to 6, wherein the demographic data comprises at least two of: age. biological sex, height, weight, and ethnicity.

8. The method of any of claims 1 to 7, wherein training the computational model further comprises training on longitudinal paired data comprising multiple image sets and corresponding reference measurements for individual training subjects captured at different time points.

9. The method of claim 8, wherein the different time points are separated by intervals of approximately four to eight weeks.

10. The method of any of claims 1 to 9, wherein training the computational model comprises using a loss function comprising a term that penalizes deviations between the sum of predicted compartment masses and the subject's measured body weight.

11. The method of any of claims 1 to 10, wherein the threshold temporal window is ten minutes or less.

12. A computer-implemented method for determining body composition of a subject, the method performed by a server system comprising one or more processors and memory storing a trained computational model, the method comprising:receiving, from a mobile computing device, a set of digital images of the subject comprising a plurality of views of the subject;preprocessing the digital images to isolate the subject from the image background; applying a trained computational model to the preprocessed image data, wherein the trained computational model performs joint inference of body composition metrics through pixel-level analysis of two-dimensional images, generating predictions directly from image features; andoutputting, as a result of the joint inference, predicted values for a plurality of body composition metrics comprising at least three physiologically distinct body composition compartments, wherein the predicted values are interdependent such that the value for any one compartment constrains the values for remaining compartments to satisfy physiological consistency relationships.

13. The method of claim 12, wherein the trained computational model was trained on paired data comprising digital images and temporally aligned body composition reference measurements from at least one medical imaging device.

14. The method of claim 13, wherein the temporal alignment comprises image capture within a threshold temporal window of reference measurement acquisition.

15. The method of claim 14, wherein the threshold temporal window is ten minutes or less.

16. The method of any of claims 12 to 15, wherein the trained computational model was trained using a loss function that penalizes violations of physiological mass-balance constraints.

17. The method of any of claims 12 to 16, wherein the plurality of views comprises an anterior view, a posterior view, and at least one lateral view.

18. The method of any of claims 12 to 17, wherein the at least three physiologically distinct body composition compartments comprise: fat mass, lean mass, and at least one additional compartment selected from the group consisting of: bone mineral content, total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

19. The method of any of claims 12 to 18, further comprising outputting regional body composition measurements for body segments selected from the group consisting of: right arm, left arm, right leg, left leg, and trunk.

20. The method of any of claims 12 to 19, further comprising receiving the subject’s measured body weight, wherein the predicted compartment values are computed as percentages of the measured body weight to satisfy the physiological consistency relationships.

21. The method of any of claims 12 to 20, wherein the predicted values for fat mass demonstrate a mean absolute error of less than five percent when evaluated against measurements from DualEnergy X-ray Absorptiometry (DEX A).

22. The method of any of claims 12 to 20, wherein the predicted values for fat mass demonstrate a mean absolute error of less than seven percent when evaluated against measurements from at least one medical imaging device.

23. The method of any of claims 12 to 22, wherein the mean absolute error for fat mass is less than four percent when validated against DEXA measurements.

24. The method of any of claims 12 to 23, wherein the trained computational model achieves prediction accuracy for body fat percentage with coefficient of determination (R2) greater than 0.65 when validated against reference measurements.

25. The method of claim 24, wherein the coefficient of determination may be greater than 0.75 for measurements obtained within ninety days of reference device measurement.

26. The method of any of claims 12 to 25, wherein the trained computational model was trained on a dataset comprising at least five hundred paired samples of digital images and reference body composition measurements.

27. The method of any of claims 12 to 26, wherein the physiological consistency relationships require that predicted compartment masses sum to a value within a predefined tolerance of the subject's measured body weight.

28. The method of any of claims 12 to 27, wherein the interdependent output variables arise from a constraint-enforcing architecture wherein:during training, the composite loss function penalizes constraint violations causing the model to leam parameter values that produce constraint-satisfying outputs; and during inference, the model produces percentage-based outputs for each compartment that are constrained to sum to one hundred percent before multiplication by measured body weight to yield absolute mass values.

29. The method of any of claims 12 to 28, wherein the body composition compartments further comprise one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

30. The method of any of claims 12 to 29, wherein the method is performed without computing any intermediate anthropometric measurements from the digital images.

31. The method of any of claims 12 to 30, wherein the method is performed through direct analysis of two-dimensional images without intermediate three-dimensional body reconstruction, thereby eliminating depth-estimation error sources and reducing computational complexity compared to volumetric reconstruction approaches while enabling multi-compartment differentiation that densitometric methods cannot achieve.

32. The method of any of claims 12 to 31, wherein inference is performed without computing body volume or body density as intermediate values.

33. The method of any of claims 12 to 32, wherein the trained computational model enforces physiological constraints during model framing through the composite loss function rather than through post-processing adjustment or temporal filtering of model outputs.

34. A computer-implemented method for tracking body composition changes in a subject over time, the method performed by a computing system comprising at least one processor and non-transitory memory, the method comprising:receiving a first set of digital images of the subject captured at a first time point; determining a first set of body composition metrics from the first set of digital images using a trained computational model;receiving a second set of digital images of the subject captured at a second time point temporally separated from the first time point by a time interval ranging from one day to three hundred sixty-five days;determining a second set of body composition metrics from the second set of digital images using the trained computational model;computing changes between the first set and the second set of body composition metrics, wherein the computed changes comprise coordinated changes across at least three physiologically distinct body composition compartments whose values remain constrained by physiological consistency relationships;computing a confidence metric for the determined body composition metrics at the second time point, wherein the confidence metric decreases as the time interval increases; andgenerating a recommendation for obtaining a reference calibration measurement when the confidence metric falls below a predetermined threshold;wherein the confidence metric indicates quantified prediction reliability.

35. The method of claim 34, wherein the prediction reliability is a function of time elapsed since reference calibration.

36. The method of any of claims 34 to 35, wherein computing the confidence metric comprises analyzing historical validation data showing prediction accuracy degradation as a function of time elapsed since reference measurement.

37. The method of any of claims 34 to 36, wherein the confidence metric comprises a mean absolute error estimate that increases with the time interval.

38. The method of any of claims 34 to 37, wherein the recommendation may be generated when the time interval exceeds ninety days.

39. The method of any of claims 34 to 38, further comprising generating a health risk indicator, wherein the health risk indicator comprises a risk assessment for at least one condition selected from the group consisting of: osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolic syndrome, and non-alcoholic fatty liver disease.

40. The method of any of claims 34 to 39, further comprising generating an alert when a computed change in at least one body composition metric exceeds a predetermined threshold.

41. The method of any of claims 34 to 40, wherein the trained computational model was trained on longitudinal data comprising paired images and medical device measurements from subjects at multiple time points, enabling the model to predict physiologically realistic body composition trajectories.

42. The method of any of claims 34 to 41 , wherein computing changes further comprises differentiating between transient changes attributable to hydration status and structural changes attributable to fat mass or lean mass modification.

43. The method of any of claims 34 to 42, wherein the trained computational model performs inference without computing total body volume, body density, anatomical keypoints, or scaled external body feature dimensions as intermediate values.

44. A system for body composition analysis comprising:a computer- readable memory storing computer-executable instructions: andone or more processors that execute the computer-executable instructions to at least: receive a set of digital images comprising a plurality of views of a subject in a standardized anatomical position;isolate the subject from the image background to generate a set of modified images;apply the modified images as input to a trained machine learning model to cause the machine learning model to generate a set of predictions for at least three physiologically distinct compartments based on jointly inferring predicted values for a plurality of body composition metrics, wherein the predicted values are generated by the trained machine learning model based on pixel-level analysis of the modified images, wherein the set of predictions are interdependent, and wherein at least a first prediction of the set of predictions constrains at least a second prediction of the set of predictions; andgenerate a health assessment based on the set of predictions.

45. The system of claim 44, further comprising a mobile computing device comprising:a second computer-readable memory storing second computer-executable instructions; anda second one or more processors that execute the second computer-executable instructions to at least:present, via a graphical user interface, instructions guiding a user to capture the set of digital images;receive the health assessment; andpresent, via the graphical user interface, the health assessment.

46. The system of claim 44, wherein the computer-executable instructions, when executed by the one or more processors, further cause the one or more processors to:generate instructions to cause display of a graphical user interface on a mobile device, wherein the graphical user interface comprises instructions for a user to generate the set of images; andtransmit the instructions to the mobile device.

47. The system of claim 45, wherein the mobile computing device provides real-time feedback to guide the user in positioning, framing, and environmental conditions during image capture.

48. The system of any of claims 44 to 47, wherein the at least three physiologically distinct compartments comprise fat mass, lean mass, and bone mineral content.

49. The system of any of claims 44 to 48, wherein the body composition metrics further comprise one or more of: total body water, visceral adipose tissue volume, and subcutaneous adipose tissue volume.

50. The system of any of claims 44 to 49, wherein the health assessment comprises diagnostic or medical pre-screening for at least one condition selected from the group consisting of: osteoporosis, osteopenia, sarcopenia, cachexia, obesity, metabolic syndrome, non-alcoholic fatty liver disease, polycystic ovary syndrome, hypothyroidism, and chronic kidney disease.

51. The system of any of claims 44 to 50, wherein the computer-executable instructions further cause the one or more processors to track the predicted body composition metrics over time and generate trend analyses and alerts based on detected changes.

52. The system of any of claims 44 to 51, wherein the trained machine learning model was trained on paired data comprising digital images captured within a temporal window corresponding to body composition reference measurements from at least one medical imaging device comprising Dual-Energy X-ray Absorptiometry (DEXA).

53. The system of any of claims 44 to 52, wherein the body composition metrics further comprise regional measurements for body segments comprising at least two of: right arm, left arm, right leg, left leg, and trunk.

54. The system of any of claims 44 to 53, wherein isolating the subject from the image background comprises removing image backgrounds while preserving subject visual information through image segmentation techniques.

55. The system of any of claims 44 to 54, wherein the one or more processors further resize isolated subject images to standardized dimensions compatible with the trained machine learning model architecture.

56. A method for generating physiologically-constrained body composition predictions over time, the method comprising:receiving a first set of digital images of a subject captured at a first time point; determining a first set of body composition metrics from the first set of digital images using a trained computational model, wherein the first set of body composition metrics comprises at least fat mass percentage and lean mass percentage;receiving a second set of digital images of the subject captured at a second time point temporally separated from the first time point by a time interval;computing a physiologically feasible range of body composition values at the second time point based on the body composition metrics determined at the first time point, the time interval, and physiological constraints on maximum rate of body composition change derived from metabolic limits;determining a second set of body composition metrics from the second set of digital images using the trained computational model;comparing the second set of body composition metrics against the physiologically feasible range; andwhen the second set of body composition metrics falls outside the physiologically feasible range, adjusting at least one of the predicted body composition values to fall within the physiologically feasible range, or a confidence metric associated with the prediction to reflect increased uncertainty.

57. The method of claim 56, wherein the physiologically feasible range expands as the time interval increases, forming a bounded region of physiologically possible body composition changes.

58. The method of any of claims 56 to 57, wherein adjusting the confidence metric comprises increasing an uncertainty bound proportional to the degree by which the predicted values exceed the physiologically feasible range.

59. The method of any of claims 56 to 58, further comprising generating a recommendation for obtaining a reference calibration measurement when the confidence metric falls below a predetermined threshold.

60. The method of any of claims 56 to 59, wherein the confidence metric is computed based on elapsed time since the subject's most recent reference imaging device measurement.

61. The method of any of claims 56 to 60, wherein the physiological constraints on maximum rate of fat mass change are derived from published metabolic equations defining minimum achievable body fat percentage.

62. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to:receive a set of digital images comprising a plurality of views of a subject in a standardized anatomical position;preprocess the digital images to isolate the subject from image backgrounds;apply a trained machine learning model to the preprocessed images to generate predicted values for at least three physiologically distinct body composition compartments, wherein the trained machine learning model was trained on paired data comprising digital images captured within a temporal window of body composition reference measurements from at least one medical imaging device, and wherein training used a loss function enforcing physiological mass-balance constraints; andoutput the predicted values, wherein the predicted values are interdependent such that predicted compartment masses sum to a value within a predefined tolerance of the subject's measured body weight.

Images