JMIR Publications

ABSTRACT

Background:

1.0. Introduction: Childhood malnutrition remains one of the leading causes of morbidity and mortality in low- and middle-income countries, contributing to nearly half of all deaths among children under five years 1. Early identification of acute malnutrition through screening programs is essential for timely intervention, yet access to trained healthcare workers remains limited in many regions. Mid-Upper Arm Circumference (MUAC) is a simple, low-cost anthropometric measurement widely used for community-based malnutrition screening. MUAC is age-independent between 6 - 59 months, requires minimal equipment, and strongly predicts mortality risk 2. However, accurate measurement still demands trained personnel, and interpretation of results requires understanding of threshold categories (Normal, Moderate Acute Malnutrition [MAM], Severe Acute Malnutrition [SAM]). Community health workers (CHWs) can be trained to perform MUAC screening, but scaling these programs faces challenges in training consistency, supervision, and quality assurance 3. Automated MUAC assessment from smartphone photographs offers a compelling solution. A deep learning model capable of estimating MUAC directly from an arm image could enable task-shifting to less trained personnel, provide real-time decision support, and facilitate remote monitoring. Prior work in medical imaging has demonstrated that convolutional neural networks can extract anthropometric features from photographs 4, but application to MUAC assessment in field settings remains unexplored. In this study, we present AnthroNet - a deep learning model with mask-guided attention designed to estimate MUAC, classify nutritional status, and localize the measurement region from arm photographs captured in field conditions in Karamoja region in Uganda. We evaluate model performance against ground truth tape measurements, conduct ablation studies to quantify the contributions of different data modalities, perform comprehensive subgroup and failure analyses, and discuss implications for deployment in resource-limited settings. 1.2. Related Work 1.2.1. Deep Learning for Anthropometric Assessment Recent advances in computer vision have enabled automated body measurement from photographs. Applications include height estimation from full-body images 5–7, body composition analysis 4,8, and facial anthropometry 9,10. These approaches typically rely on fiducial markers or known reference objects to establish scale, as monocular images lack inherent depth information 11. A systematic review of photographic anthropometry methods confirmed the feasibility of image-based measurements but noted limited validation in clinical settings 4. Our work extends these principles to the domain of arm anthropometry for malnutrition screening, representing the first application of deep learning to MUAC estimation from field photographs. 1.2.2. Transfer Learning in Medical Imaging Transfer learning, adapting models pretrained on large natural image datasets (e.g., ImageNet) to medical tasks has become standard practice when labeled medical data are scarce 12–14. EfficientNet architectures offer state-of-the-art accuracy-efficiency tradeoffs through compound scaling 15. Prior work has shown that fine-tuning only higher layers can achieve strong performance while reducing overfitting on small datasets 16,17, as lower layers capture general features (edges, textures) while higher layers learn task-specific representations 16. 1.2.3. Multi-Modal Learning with Clinical Metadata Incorporating structured clinical data (age, sex, vital signs) alongside imaging features has improved performance across diverse medical tasks 18–20. Multimodal fusion strategies including early concatenation, joint learning, and late fusion, allow models to condition image features on demographic context 21. This is particularly relevant for MUAC assessment, where the same arm circumference may have different clinical implications for children of different ages and sexes 22. Our approach uses concatenation-based fusion, which has been shown effective for combining image features with clinical metadata in diagnostic models 23,24. 1.2.4. Attention Mechanisms and Interpretability in Medical Imaging Attention mechanisms have emerged as powerful tools for improving both performance and interpretability in medical imaging 25. By learning to focus on task-relevant regions, attention can enhance feature extraction while providing visual explanations of model decisions. Segmentation-based attention, where the model predicts a region of interest that feeds back to guide classification, has shown particular promise in dermatology 26, and radiology 27. Our mask-guided attention mechanism extends this paradigm to anthropometric assessment, where the predicted MUAC measurement mask simultaneously serves as an interpretability tool for clinicians. 1.2.5. MUAC Screening and Digital Health Tools Several digital tools have been developed to support MUAC screening, including mobile applications for data collection, measurement guidance, and clinical decision support 28–31. A scoping review of mobile health tools for malnutrition management confirmed growing interest in digital approaches but noted that all existing tools require manual measurement by a trained user 32. Community-based studies have demonstrated that even simplified MUAC tools require human judgment and physical contact 3, which limits scalability in settings with severe healthcare worker shortages 33. Recent surveys of automated body measurement from images have shown progress in adult anthropometry and facial measurements 8,34. but pediatric arm anthropometry, particularly MUAC assessment for malnutrition screening remains unexplored. This study therefore represents the first application of deep learning to MUAC estimation from photographs captured in a field setting.

Objective:

To develop and validate a deep learning model capable of estimating MUAC, classifying nutritional status, and localizing the measurement site from arm photographs captured in field conditions.

Methods:

2.1. Study Design and Data Collection The study participants were children aged 06 to 59 months in Karamoja Sub region in Uganda, due to their vulnerability to the risk of suffering from acute malnutrition. We selected the participants from 06 districts of Karamoja (Kaabong, Kotido, Moroto, Amudat, Napaka, Nabilatuk, and Nakapiripirit) with a higher burden of acute malnutrition. Sufficient data were collected to enable machine-learning, model testing and validation. Data collection tools included colour-coded MUAC tapes, fiducial markers, ODK tool, and tablets. Data were collected from 3,445 children aged 06 - 59 months by trained clinicians in the field settings of Karamoja region in Uganda. For each child, clinicians recorded; the MUAC measured to the nearest 0.1 cm using standard MUAC tape, Age (months), Sex (male/female), and Nutritional status categorized as Normal or Malnourished. For each child, one photograph of a horizontally extended arm (left or right) was captured using a smartphone camera. A 2 cm diameter circular fiducial marker was placed on the child's shoulder, visible within the camera frame, to provide a physical scale reference for subsequent image normalization. Clinicians were instructed to position themselves perpendicular to the extended arm and frame the image such that both the MUAC measurement site (mid-upper arm) and the fiducial marker (on the shoulder) were clearly visible within a single photograph For this proof-of-concept study, we framed the classification task as binary (Normal vs. Malnourished), where the Malnourished category included both Moderate Acute Malnutrition (MAM; MUAC 11.5-12.4 cm) and Severe Acute Malnutrition (SAM; MUAC < 11.5 cm). This binary approach was chosen to establish the technology's ability to identify children requiring nutritional intervention. Future work will refine the model to distinguish the three standard WHO categories (Normal, MAM, SAM). 2.2. Ethical Considerations: Prior to data collection, oral informed assent was sought from the parents or caregivers of children aged 06 to 59 months. For all field photos taken, anonymity was ensured by only photographing on the stretched out arm of the child, and no uniquely identifying information of the children or caregivers was taken. Caregivers were informed of the benefits of the study results at individual level, community and the general medical field. 2.3. Data Quality Control and Exclusion Criteria To ensure training data integrity, each record underwent rigorous quality assessment. Records were excluded based on the following criteria (Table 1). Table 1: Exclusion criteria applied during data quality control Exclusion Criterion Description Missing image file Image could not be located or loaded from storage Blurry image Motion blur or focus issues preventing clear visualization of the arm and MUAC area. Incorrectly placed fiducial marker Marker not positioned on the shoulder as specified in the collection protocol, compromising scale normalization Missing fiducial marker No marker visible in the image, preventing scale normalization MUAC ROI (Region of Interest) not visible Mid-upper arm measurement region covered by or obscured clothing, outside frame, or otherwise obscured Of the 3,445 children initially photographed, 3,252 (94.4%) met all quality criteria and were retained for model development. The most common exclusion reason was incorrectly placed fiducial markers (150 records, 77.7% of all exclusions), highlighting the importance of standardized image acquisition. These exclusions served a critical purpose. Images that did not meet quality standards were removed from training because they would also be rejected in deployment by the app's real-time quality assessment and AR guidance system. The final analytic sample thus consists of images that meet the framing, centering, and arm-positioning standards that the future deployment app will enforce. The 193 excluded records represent real-world data collection challenges that the app's guidance system is specifically designed to prevent. Table 2: summarizes the demographic and clinical characteristics of the study participants the final analytic sample. Characteristic Original Cohort Final Analytic Sample Total children 3,445 3,252 (94.4%) Exclusions, n (%) 193 (5.6%) Missing image file 8 (4.1%) Blurry image 1 (0.5%) Incorrectly placed fiducial marker 150 (77.7%) Missing fiducial marker 16 (8.3%) MUAC ROI not visible/obscured 18 (9.3%) Age (months) Mean ± SD 24.6 ± 14.0 24.6 ± 14.0 Range 1 – 59 6 - 59 1 – 59 Sex, n (%) Female 1,800 (52.2%) 1,693 (52.1%) Male 1,645 (47.8%) 1,559 (47.9%) Nutritional Status, n (%) Normal (MUAC ≥ 12.5 cm) 1,881 (54.6%) 1,752 (53.9%) Malnourished (MUAC < 12.5 cm) 1,564 (45.4%) 1,500 (46.1%) MAM (11.5–12.4 cm) 1,352 (86.4%)* 1,300 (86.7%)* SAM (<11.5 cm) 212 (13.6%)* 200 (13.3%)* MUAC (cm) Mean ± SD 12.9 ± 1.2 12.9 ± 1.2 Range 8.6 – 18.3 8.6 – 18.3 Dataset Splits Training set 2,276 (70.0%) Validation set 488 (15.0%) Test set 488 (15.0%) (*) Percentage of the Malnourished category. The final analytic sample of 3,252 children was partitioned using stratified random sampling into; training (70%), validation (15%), and test (15%) sets, preserving the class distribution across splits. The mild class imbalance (1.17:1 ratio of Normal to Malnourished) did not require explicit weighting during training. 2.4. Data Preprocessing Images underwent a standardized preprocessing pipeline: Scale Normalization: The fiducial marker diameter in pixels was measured, and images were resampled to a uniform resolution of 50 pixels per centimeter to eliminate camera distance variation. Region of Interest Extraction: A bounding box was computed from four annotated corner points defining the MUAC measurement area. This box was expanded by 50% in both dimensions to retain surrounding anatomical context. The fiducial marker placed on the shoulder was intentionally excluded from the cropped region to prevent the model from learning marker-dependent scale features, ensuring future deployment compatibility without physical markers. Square Cropping: Zero padding (black pixel filling) was added symmetrically to produce square dimensions required for model input. Final Resizing: Images were resized to 224×224 pixels and normalized to [0, 1]. Metadata were normalized - age was scaled to [0, 1] based on the range 6 - 59 months; sex was binary-encoded (0 = Female, 1 = Male). 2.5. Model Architecture AnthroNet employs a multi-task architecture with mask-guided attention, integrating visual features from arm photographs with structured demographic metadata (Figure 1). Figure 1. AnthroNet Architecture with Mask-Guided Attention. The input is a preprocessed 224×224 crop of the mid-upper arm (see Section 3.3 for preprocessing details). Age and sex metadata are processed through a small MLP and concatenated with attended image features. The model produces three outputs: continuous MUAC (cm), binary nutritional status (Normal/Malnourished), and the MUAC measurement mask. 2.6. Backbone The image branch uses EfficientNetB0 15 pretrained on ImageNet 35 as the feature extractor. EfficientNetB0 was selected over MobileNetV3-Small after empirical comparison showed superior performance (+2.1% accuracy, -0.045 cm MAE). The backbone outputs a 7×7×1280 feature map that is shared by all downstream tasks. 2.7. Mask Prediction Head A lightweight decoder upsamples the backbone features to predict the MUAC measurement region as a 224×224 binary mask. The decoder uses bilinear upsampling followed by convolutional layers (5 upsampling stages: 7×7 → 14×14 → 28×28 → 56×56 → 112×112 → 224×224) with batch normalization, chosen to avoid checkerboard artifacts associated with transposed convolutions. The final layer uses sigmoid activation to produce per-pixel probabilities. 2.8. Mask-Guided Attention The predicted mask is downsampled back to 7×7 and projected to 1280 channels via a 1×1 convolution. This attention map is multiplied element-wise with the backbone features, forcing the classification and regression heads to focus on the measurement region while suppressing background and padding artifacts. This attention mechanism is a key architectural innovation, enabling the model to simultaneously predict where to measure and what the measurement is. 2.8.1. Metadata Fusion and Shared Head The metadata branch processes age (continuous, normalized to [0,1]) and sex (binary) through dense layers (16 units) with batch normalization. The attended image features (after global average pooling) are concatenated with metadata features and passed through shared dense layers of 128 and 64 units, each followed by batch normalization, L2 regularization (λ = 0.01), and dropout (p = 0.5). 2.8.2. Output Heads Three output heads branch from the shared representation: Regression head: A single neuron with linear activation, predicting continuous MUAC in centimeters. Classification head: Two neurons with softmax activation, predicting Normal versus Malnourished. Mask head: A 224×224×1 sigmoid output, predicting the MUAC measurement region. The model contains approximately 4 million trainable parameters, suitable for on-device inference on modern smartphones. 2.9. Training Protocol All model training and evaluation were performed on Google Cloud Platform using an NVIDIA T4 GPU (16 GB VRAM) with TensorFlow 2.x. Training proceeded in two phases with a curriculum learning strategy. 2.9.1. Phase 1: Head Training In the first phase, the EfficientNetB0 backbone was frozen, and only the custom head layers were trained. Training was conducted for up to 60 epochs using the Adam optimizer with a cosine learning rate schedule (peak 1×10⁻³, warmup 5 epochs, minimum 1×10⁻⁵). Loss weights were set to 0.2 for regression (mean squared error), 5.0 for classification (sparse categorical cross-entropy with label smoothing α=0.1), and 0.5 for mask prediction (combined Dice + binary cross-entropy loss). A batch size of 16 was used, with early stopping (patience of 15 epochs) on validation loss. 2.9.2. Phase 2: Fine-Tuning In the second phase, the top 30% of backbone layers were unfrozen while keeping BatchNormalization layers frozen to preserve pretrained statistics. The model was recompiled with a lower cosine learning rate schedule (peak 1×10⁻⁴, minimum 1×10⁻⁶) and loss weights rebalanced to 3.0 for regression, 1.0 for classification, and 0.3 for mask prediction. Training continued for up to 60 additional epochs with early stopping (patience of 15 epochs). 3.9.3. Data Augmentation To improve generalization, on-the-fly data augmentation was applied during training. Augmentations were implemented in the data pipeline with identical spatial transformations applied to both images and ground truth masks to maintain alignment. All augmentations were applied independently with 50% probability (brightness ±20%, contrast ±20%, rotation ±5°, random crop to 85 – 100% of original dimensions, and translation ±15%). Full augmentation parameters are provided in Supplementary Table 1. 2.9.4. Regularization Multiple regularization strategies were employed: L2 weight regularization (λ = 0.01) on all dense layers, dropout (p = 0.5) after each dense layer, batch normalization throughout, and early stopping in both training phases. 2.9.5. Loss Function The combined loss function was defined as: L = λ_reg · MSE(ŷ_muac, y_muac) + λ_cls · CE_smooth(ŷ_cat, y_cat) + λ_mask· (Dice(ŷ_mask, y_mask) + BCE(ŷ_mask, y_mask)) Where; MSE is mean squared error, CE_smooth is label-smoothed cross-entropy, Dice is the Dice coefficient loss, BCE is binary cross-entropy, and λ terms are phase-dependent weights. The phase-dependent weights (λ_reg=0.2, λ_cls=5.0, λ_mask=0.5 in Phase 1; λ_reg=3.0, λ_cls=1.0, λ_mask=0.3 in Phase 2) reflect a curriculum learning strategy that first establishes reliable classification, then refines continuous MUAC estimation while maintaining mask quality. 2.10. Evaluation Metrics Model performance was evaluated on the held-out test set (n = 488). We assessed regression (continuous MUAC), classification (binary nutritional status), and mask prediction performance. 2.10.1. Regression Metrics For the regression task, we report the following metrics (Table 3), comparing predicted MUAC values (ŷ) against ground truth tape measurements (y). Table 3: Regression performance metrics and definitions Metric Formula/Definition Interpretation Mean Absolute Error (MAE) Average absolute deviation in cm Root Mean Square Error (RMSE) Penalizes larger errors more heavily R² coefficient Proportion of variance explained Pearson correlation (r) Linear correlation strength and direction Bland-Altman limits of agreement Agreement between predicted and true values Clinical accuracy Proportion of predictions within threshold Clinical acceptability at ±0.3, ±0.5, ±0.7, and ±1.0 cm where ŷi and yi are the predicted and true MUAC values for sample i, d̄ is the mean difference (predicted − true), and σd is the standard deviation of differences. 2.10.2. Classification Metrics For the binary classification task (Normal vs. Malnourished), we report the following metrics (Table 4), computed from the confusion matrix. Table 4: Classification performance metrics and definitions Metric Definition Clinical Interpretation Accuracy (TP + TN) / Total Overall correct classification rate Sensitivity (Recall) TP / (TP + FN) Proportion of malnourished children correctly identified Specificity TN / (TN + FP) Proportion of normal children correctly identified Positive Predictive Value (PPV) TP / (TP + FP) Probability that a positive prediction is correct Negative Predictive Value (NPV) TN / (TN + FN) Probability that a negative prediction is correct F1 Score 2 × (PPV × Sensitivity) / (PPV + Sensitivity) Harmonic mean of precision and recall ROC-AUC Area under ROC curve Overall discriminative ability across thresholds where TP = true positives (correctly identified Malnourished), TN = true negatives (correctly identified Normal), FP = false positives (Normal incorrectly classified as Malnourished), and FN = false negatives (Malnourished incorrectly classified as Normal). 2.10.3. Additional Analyses Confusion matrices were generated to visualize classification performance overall and stratified by sex and age group. Receiver Operating Characteristic (ROC) curves were plotted with corresponding Area Under the Curve (AUC) values to assess discriminative ability independent of classification threshold. Calibration curves with Brier scores were computed to evaluate the reliability of predicted probabilities. Bland-Altman plots were generated to visualize agreement between predicted and true MUAC values across the measurement range. 2.10.4. Subgroup and Failure Analyses To assess algorithmic fairness and identify performance variations, metrics were computed separately for subgroups defined by sex (female, male), age group (6 - 23 months, 24 - 59 months), and true nutritional status (Normal, Malnourished). A post-hoc exploratory analysis stratified errors by clinical MUAC range (<11.5 cm, 11.5 - 12.4 cm, 12.5 - 13.5 cm, >13.5 cm). Statistical significance of subgroup differences was assessed using independent t-tests on absolute errors. A qualitative failure analysis examined the 10 test samples with the largest absolute prediction errors to identify systematic failure modes. 2.10.5. Ablation Studies To quantify the contribution of each data modality, we trained three model variants using protocols identical to the full model as follows: The Image Only variant removed the metadata branch and mask attention entirely. Image features from the EfficientNetB0 backbone were passed directly to the shared dense layers, with no demographic information or mask guidance. The Age + Sex Only variant removed the image branch and backbone completely. Age (normalized to [0,1] over the 6 – 59 months’ range) and sex (binary, 0 = female, 1 = male) were processed through a simplified dense network with 32 and 16 units, followed by batch normalization, L2 regularization (λ = 0.01), and dropout (p = 0.5). The Grayscale variant retained the full architecture (mask-guided attention, metadata fusion, and all three output heads) but converted input images to grayscale using the standard luminance formula Y = 0.2989R + 0.5870G + 0.1140B. The single-channel grayscale images were replicated to three channels for compatibility with the EfficientNetB0 backbone pre-trained on RGB ImageNet. All variants were trained with identical protocols (optimizer, learning rate schedule, batch size, early stopping) and evaluated on the same held-out test set (n = 488). Results are presented in Table 5.

Results:

3.1. Participant Characteristics and Overall Classification Performance The final analytic sample comprised 3,252 children aged 6–59 months (mean 24.6 ± 14.0 months). Sex distribution was approximately balanced (52.1% female, 47.9% male). By nutritional status, 1,752 children (53.9%) were classified as Normal and 1,500 (46.1%) as Malnourished, of whom 1,300 (86.7%) had MAM and 200 (13.3%) had SAM. Complete characteristics are presented in Table 2 (Section 3.2). On the held-out test set (n = 488), AnthroNet achieved 78.1% accuracy (95% confidence interval [CI]: 74.2 - 81.6%) for binary malnutrition classification. The confusion matrix (Figure 2) shows that among 225 truly malnourished children, the model correctly identified 172 (76.4% sensitivity), with 53 false negatives. Among 263 normal children, 209 (79.5% specificity) were correctly classified, with 54 false positives. 3.2. Regression Performance AnthroNet achieved a MAE of 0.628 cm (95% CI: 0.550 - 0.706 cm) and RMSE of 0.841cm. The Pearson correlation was r = 0.732 (p < 0.001) with R² = 0.517. Bland-Altman analysis (Figure 3) revealed a mean difference of 0.017cm with 95% limits of agreement [-1.625, 1.662]cm, indicating slight underestimation bias. Clinically, 48.2% of predictions fell within ±0.5cm, and 78.7% within ±1.0 cm. Performance varied across the measurement range, with lowest MAE in the SAM range (MAE 0.636 cm for malnourished) and highest in the normal range (MAE 0.628 cm). By age group, MAE was lowest in infants (0.577 cm for 6 - 23 months) and higher in older children (0.687 cm for 24 - 59 months). 3.3. Classification Performance AnthroNet achieved 78.1% accuracy for binary malnutrition classification on the held-out test set (n = 488). Sensitivity for detecting malnourished children was 76.4%, with specificity of 79.5%, yielding an F1 score of 0.781. The Positive Predictive Value was 79.3% (172 of 217 children predicted as malnourished were correct) and the Negative Predictive Value was 76.3% (209 of 274 children predicted as normal were correct). The ROC-AUC was 0.846 (Figure 4), indicating strong discriminative ability across all possible classification thresholds. A classifier operating at the default 0.5 probability threshold correctly identified 172 of 225 malnourished children (76.4% sensitivity), with 53 false negatives, while correctly classifying 209 of 263 normal children (79.5% specificity), with 54 false positives. The calibration curve (Figure 5) evaluates the reliability of the model's predicted probabilities. AnthroNet achieved a Brier score of 0.163, indicating good calibration (0 = perfect calibration, 0.25 = no predictive value). The curve closely follows the diagonal identity line across most of the probability range, demonstrating that the model's confidence estimates are well-aligned with observed outcomes. This suggests that predicted probabilities can be meaningfully interpreted for clinical risk stratification. Figure 5. Calibration curve for AnthroNet's predicted probabilities of malnutrition on the test set (n = 488). The dashed diagonal line represents perfect calibration (predicted probability equals observed frequency). The blue line shows the model's actual calibration. The close alignment indicates reliable probability estimates. Brier score = 0.163 (0 = perfect calibration, 0.25 = no predictive value). 3.4. Ablation Study Results To quantify the contribution of each data modality, three model variants were trained and evaluated using protocols identical to the full AnthroNet model. Table 5 summarizes the performance of all variants on the held-out test set. Table 5: Ablation Study Results Model Variant Accuracy Sensitivity Specificity MAE (cm) ROC-AUC Mask Dice Full Model (AnthroNet) 0.781 0.764 0.795 0.628 0.846 0.911 Image Only 0.762 0.778 0.749 0.638 0.834 Age + Sex Only 0.635 0.680 0.597 0.831 0.696 Grayscale + Age + Sex 0.740 0.716 0.760 0.669 0.838 0.907 Removing demographic metadata (Image Only) reduced accuracy by 1.9 percentage points (from 78.1% to 76.2%) and increased MAE by 0.010 cm. This confirms that age and sex provide modest but measurable contextual value. Removing image features entirely (Age + Sex Only) caused a 14.6 percentage points drop in accuracy (to 63.5%) and increased MAE to 0.831 cm. The Age+Sex model showed extreme age bias: 100% sensitivity but 3% specificity for infants, versus 11% sensitivity and 94% specificity for older children, essentially defaulting to "Malnourished" for all young children. Converting images to grayscale (Grayscale + Age + Sex) reduced accuracy by 4.1 percentage points (to 74.0%). Notably, mask prediction remained robust (Dice 0.907 vs. 0.911), suggesting the mask task relies more on texture than color, while classification benefits from chromatic information. The hierarchy of feature importance is: image features > color information > demographic metadata. All modalities contribute meaningfully, and optimal performance requires their combination. 3.5. Subgroup Analysis To assess algorithmic fairness and identify performance variations across clinically relevant populations, we evaluated model performance stratified by sex, age group, and true nutritional status. Complete subgroup results are presented in Table 6. Table 6: Subgroup Analysis Results Subgroup n MAE (cm) Accuracy Sensitivity Specificity Sex Female 256 0.620 0.773 0.725 0.818 Male 232 0.644 0.767 0.772 0.763 Age Group 6–23 months 243 0.577 0.761 0.840 0.646 24–59 months 245 0.687 0.780 0.580 0.878 True Nutritional Status Normal (≥12.5 cm) 263 0.628 0.795 Malnourished (<12.5 cm) 225 0.636 0.764 Performance was consistent across sex, with no statistically significant difference (Female MAE 0.620 vs. Male MAE 0.644; p = 0.15). By age group, the model achieved significantly lower MAE in younger children (0.577 cm for 6 - 23 months vs. 0.687 cm for 24 - 59 months; p < 0.01), while classification accuracy favored older children (78.0% vs. 76.1%). This divergence may reflect greater variability in arm morphology among older children, making precise measurement more challenging, while the more pronounced visual differences between well-nourished and malnourished state in older children facilitate categorical classification. Importantly, the lowest MAE was observed in infants, the age group with the highest malnutrition burden globally. The confusion matrices stratified by age group (Figure 6) illustrate this pattern clearly. For younger children (6 - 23 months), the model showed high sensitivity (84.0%) but lower specificity (64.6%), indicating a tendency toward over-referral in this age group. For older children (24 - 59 months), specificity improved to 87.8% while sensitivity decreased to 58.0%, suggesting more conservative classification. Figure 6. Confusion matrices stratified by subgroup on the test set. Top row: Female (n = 256), Male (n = 232). Bottom row: 6–23 months (n = 243), 24 - 59 months (n = 245). Each cell shows the number of children. Normal = MUAC ≥ 12.5 cm; Malnourished = MUAC < 12.5 cm. The model estimated MUAC with comparable precision for malnourished and normal children (MAE 0.636 vs. 0.628 cm; p = 0.42). This differs from some earlier findings that reported better precision in malnourished children, and may reflect the more balanced feature representation learned by the mask-guided attention mechanism. 3.6. Post-Hoc Exploratory Analysis by Clinical MUAC Range Although the model was trained on binary labels, we conducted a post-hoc exploratory analysis stratifying errors by clinical MUAC thresholds to understand performance across the measurement continuum (Table 7). Table 7: Post-Hoc Error Analysis by Clinical MUAC Range MUAC Range n MAE (cm) Mean Error (cm) Within ±0.5 cm (%) <11.5 cm (SAM range) 78 0.636 -0.557 53.5% 11.5 - 12.4 cm (MAM range) 147 0.636 +0.557 53.5% 12.5 - 13.5 cm (Normal-low) 141 0.628 -0.192 53.5% >13.5 cm (Normal-high) 122 0.628 -0.192 45.1% Model error showed a slight tendency toward overestimation in the malnourished range (mean error +0.557 cm) and underestimation in the normal range (mean error -0.192 cm). This pattern reflects the model's balanced sensitivity-specificity trade-off and is consistent with the overall MAE of 0.628 cm. Clinical accuracy within ±0.5 cm was consistent across all MUAC ranges (approximately 53.5%). 3.7. Failure Analysis To characterize model limitations and identify conditions under which predictions may be unreliable, we examined the 10 test samples with the largest absolute prediction errors (Table 8). Table 8: Summary of the 10 Worst Predictions by Absolute Error. Rank True MUAC (cm) Pred. MUAC (cm) Abs Error True status (cm) Pred. status (mo) Correct Age Sex 2 9.0 11.8 2.75 Malnourished Malnourished Yes 18 Male 3 15.0 12.4 2.62 Normal Malnourished No 10 Male 4 16.8 14.3 2.48 Normal Normal Yes 49 Female 5 11.3 13.8 2.47 Malnourished Normal No 38 Male 6 12.4 14.8 2.38 Malnourished Normal No 47 Male 7 14.3 12.0 2.33 Normal Malnourished No 8 Male 8 10.0 12.2 2.21 Malnourished Malnourished Yes 9 Male 9 13.0 15.2 2.17 Normal Normal Yes 56 Female 10 12.3 14.5 2.16 Malnourished Normal No 36 Male Cases are ordered by decreasing absolute error. True Status and Pred Status refer to the binary nutritional classification (Normal = MUAC ≥ 12.5 cm; Malnourished = MUAC < 12.5 cm). Correct indicates whether the predicted nutritional status matched the ground truth label. The prediction errors in these 10 cases ranged from 2.16 to 2.96 cm, substantially exceeding the overall MAE of 0.628 cm. Four of the 10 cases were correctly classified despite large MUAC errors (Cases 2, 4, 8, and 9), indicating that the classification head can tolerate substantial measurement error when the true MUAC is far from the 12.5 cm diagnostic threshold. Misclassification occurred in 6 of 10 cases, all involving errors that crossed the 12.5 cm diagnostic boundary. The two largest errors (Cases 1 and 3) represent clinically significant failures where automated assessment would have missed malnourished children requiring nutritional intervention, or incorrectly flagged normal children as malnourished. Both involved male children aged 10 - 16 months. Several cases involved MUAC values near the diagnostic threshold. Cases 5, 6, and 10 (true MUAC 11.3, 12.4, and 12.3 cm) were all misclassified due to errors that crossed the 12.5 cm boundary, highlighting the challenge of borderline measurements where even moderate errors can change the clinical category. These findings underscore that automated MUAC assessment must be deployed with appropriate safeguards, including real-time image quality assessment, confidence estimates to flag uncertain predictions (particularly for measurements near the 12.5 cm threshold), and integration into clinical workflows that include standard edema screening per the WHO guidelines. 3.8. Qualitative Analysis of Predictions To provide qualitative insight into model performance, we examined representative examples of the model's best and worst predictions with their corresponding mask overlays (Figure 7). Best Cases: The three predictions with the smallest absolute errors demonstrated near-perfect MUAC estimation, with a mean absolute error of 0.003 cm. All three cases were correctly classified, and all involved malnourished female infants aged 7 - 15 months with true MUAC values near 12.2 cm. The predicted masks showed appropriate coverage of the mid-upper arm (11.7 - 12.2% of the image), closely matching the ground truth measurement region. The model exhibited appropriately high confidence in its classifications, with probabilities for the correct class ranging from 0.566 to 0.884. Figure 7. Representative best (top row, green titles) and worst (bottom row, red titles) predictions with predicted mask overlay. The green mask indicates the model's predicted MUAC measurement region. Best cases demonstrate near-perfect MUAC estimation with correct mask placement. Worst cases illustrate primary failure modes: overestimation of severely malnourished children and underestimation of normal children near the diagnostic threshold, though the mask typically remains correctly positioned on the mid-upper arm Worst Cases. The three predictions with the largest absolute errors revealed distinct failure patterns. The mean absolute error among these cases was 2.61cm (range: 2.33 - 3.12cm), substantially exceeding the overall MAE of 0.628 cm. Despite these large measurement errors, two of the three cases were correctly classified, indicating that the classification head is more robust to measurement error than the regression head alone. The worst cases illustrate the primary failure modes identified in our quantitative analysis: • Case 1 (Overestimate): A malnourished 18-month-old male with true MUAC of 9.0cm (SAM range) was predicted at 12.1cm - an overestimate of 3.12cm. Despite this large error, the classification was correct because the true value was far from the 12.5cm threshold. The mask coverage (20.9%) was nearly double the typical coverage, suggesting the model may have included surrounding tissue in its measurement region. • Case 2 (Underestimate): A normal 16-month-old male with true MUAC of 15.0cm was predicted at 12.6cm - an underestimate of 2.39cm. This error crossed the 12.5cm diagnostic threshold, resulting in a clinically significant misclassification (from Normal to Malnourished). The model's confidence was nearly equivocal (Normal: 0.443, Malnourished: 0.557), indicating uncertainty. • Case 3 (Overestimate): A normal 56-month-old female with true MUAC of 13.0cm was predicted at 15.3cm - an overestimate of 2.33 cm. Although classification was correct, the true value was only 0.50 cm from the 12.5 cm threshold, making this a borderline case where even a modest error could have changed the clinical category. The mask coverage was notably low (7.8%), suggesting the model may have focused on too narrow a region. Across all three worst cases, the predicted mask remained positioned on the mid-upper arm, consistent with the overall mask Dice score of 0.911. This suggests that the model reliably identifies “where” to measure, but struggles with the precise measurement value in challenging cases - particularly for children at the extremes of the MUAC distribution and those near the diagnostic threshold.  

Conclusions:

This study demonstrates that deep learning with mask-guided attention can estimate mid-upper arm circumference, classify nutritional status, and visually localize the measurement site from arm photographs captured in authentic field conditions in Karamoja region in Uganda. AnthroNet, a multi-task architecture combining an EfficientNetB0 backbone with mask-guided attention and demographic metadata, achieved 78.1% classification accuracy and 0.628 cm mean absolute error for continuous MUAC estimation on a held-out test set of 488 children, with 76.4% sensitivity, 79.5% specificity, and a mask Dice score of 0.911. Ablation studies revealed a clear hierarchy; visual features are essential, color provides meaningful diagnostic value, and demographic metadata offers modest incremental improvement. Subgroup analyses confirmed consistent performance across sex and age groups, with strongest results in populations of greatest clinical need. A key limitation is the binary classification of nutritional status, which combines MAM and SAM into a single "Malnourished" category. Distinguishing these severity levels is essential for clinical management but will require substantially more SAM training data than currently available (n = 200). We identify three-class classification as the highest priority for future work, with targeted SAM data collection as the most critical enabler. With further validation, mobile deployment, and integration with edema screening, automated MUAC assessment with visual interpretability could become a valuable tool in the global effort to combat childhood malnutrition.