Feed-O-Meter: Feedback & Intake Monitoring
- Feed-O-Meter is a cross-domain systems concept that spans LLM-based design feedback tools and sensor-driven food intake monitors, providing actionable insights.
- It employs AI with audio, video, and inertial sensors to estimate bite weight, meal intervals, and nutrient metrics with precision and real-time feedback.
- Feed-O-Meter systems also enable closed-loop interventions and behavioral regulation, enhancing performance in design education, human eating, and aquaculture applications.
Searching arXiv for papers directly relevant to “Feed-O-Meter” and adjacent sensing systems. “Feed-O-Meter” denotes several distinct but conceptually related systems in recent arXiv literature. In the narrowest sense, it names an LLM-based role-playing environment for practicing design feedback through interaction with an AI mentee (Lim et al., 9 Sep 2025). In a broader technical sense suggested by adjacent work, it also functions as an organizing label for systems that meter feeding, eating, or intake-related processes through sensing, inference, and feedback: commodity-earbud bite-weight estimation (Papapanagiotou et al., 2021), smartwatch-based bite and meal monitoring (Kyritsis et al., 2020), earbud-based closed-loop eating-pace regulation (Fang et al., 10 Feb 2026), egocentric food-volume estimation (Wang et al., 7 May 2025), multimodal macronutrient estimation (Arefeen et al., 3 Mar 2025), family eating-context monitoring (Mondol et al., 2020), and precision feeding in aquaculture (Hossam et al., 2024, Georgopoulou et al., 2024). Across these usages, the recurring technical motif is a system that translates raw observations into actionable estimates about consumption, feeding state, or feedback quality.
1. Terminological scope and principal senses
The term “Feed-O-Meter” appears explicitly as the title of “Feed-O-Meter: Fostering Design Feedback Skills through Role-playing Interactions with AI Mentee” (Lim et al., 9 Sep 2025). In that work, “feed” refers to feedback rather than food or feed dispensing. The system is designed to help design students practice giving feedback in a low-pressure, interactive setting, with users role-playing as mentors to an AI “mentee” named Alex (Lim et al., 9 Sep 2025).
The same label is also used in the supplied literature as an interpretive umbrella for intake-monitoring and feeding-control systems. In “Bite-Weight Estimation Using Commercial Ear Buds,” the work is framed as addressing a practical eating-behavior sensing problem aimed at a “Feed-O-Meter-style system,” where the objective is not merely to detect eating, but to estimate the weight of a single bite using only audio captured by commercial earbuds (Papapanagiotou et al., 2021). In “Earinter: A Closed-Loop System for Eating Pace Regulation with Just-in-Time Intervention Using Commodity Earbuds,” the system is described as “essentially a practical realization of a ‘Feed-O-Meter’” because it continuously measures eating pace, decides when pace is drifting into a risky fast-eating zone, and regulates behavior with just-in-time feedback during the meal (Fang et al., 10 Feb 2026).
A plausible implication is that “Feed-O-Meter” is best understood not as a single standardized device class, but as a cross-domain systems concept. Depending on context, it may denote a pedagogical feedback meter, a human food-intake meter, an animal feeding-state monitor, or a feed-allocation controller.
2. Human eating and intake metering
In human sensing, the literature associated with a Feed-O-Meter-style interpretation centers on turning passive observations into bite-level, meal-level, or nutrient-level estimates. “Bite-Weight Estimation Using Commercial Ear Buds” evaluates the potential for estimating the weight of consumed food per bite using only the audio signal captured by Samsung Galaxy Buds, rather than specialized sensing hardware (Papapanagiotou et al., 2021). The in-house dataset includes 8 participants, 4 food types, about 2 hours of eating, 473 chewing bouts, and 7.539 chews, with food-specific models achieving mean absolute error of less than 1 g for 3 out of 4 food types and an all-food model reaching 2.12 g (Papapanagiotou et al., 2021). The work explicitly positions this as a step toward a passive intake-monitoring pipeline that could estimate bite weight and then accumulate bite estimates into meal-level tracking.
“A Data Driven End-to-end Approach for In-the-wild Monitoring of Eating Behavior Using Smartwatches” addresses a complementary problem: temporal localization rather than intake mass estimation (Kyritsis et al., 2020). Using commercially available smartwatch inertial sensors, the framework performs both in-meal bite detection and temporal localization of meals in unconstrained daily life (Kyritsis et al., 2020). The end-to-end CNN + LSTM bite detector reaches an F1 score of 0.923 under leave-one-subject-out evaluation, and the meal-localization stage achieves a Jaccard Index of 0.820 in LOSO and 0.821 in held-out evaluation (Kyritsis et al., 2020). This establishes a Feed-O-Meter-like capability for continuous monitoring of when meals start and end, even though it does not estimate the mass or nutrient composition of what is eaten.
“FoodTrack: Estimating Handheld Food Portions with Egocentric Video” advances a more geometric notion of consumption metering by estimating food volume directly from egocentric video (Wang et al., 7 May 2025). The pipeline uses Project Aria glasses, ResShift for super-resolution, Grounded SAM and Cutie for segmentation, ChronoDepth for temporal depth estimation, a modified BundleSDF pipeline with LightGlue, and mesh postprocessing with PyVista fill_holes and Trimesh’s mesh.volume (Wang et al., 7 May 2025). In a handheld sandwich example, the system estimates approximately 345 mL versus a measured 371 ± 1 mL, corresponding to an absolute percentage error of 7.01% (Wang et al., 7 May 2025). This differs from bite-based systems by reconstructing remaining or current object volume rather than inferring intake from gestures or chewing.
“MealMeter: Using Multimodal Sensing and Machine Learning for Automatically Estimating Nutrition Intake” frames the meter at the macronutrient level rather than the event or object level (Arefeen et al., 3 Mar 2025). Using Dexcom G6 CGM data, Empatica E4 signals, feature engineering, PCA with 3 principal components, and linear regression, the system predicts carbohydrates, proteins, and fats from a 90-minute post-meal physiological window plus pre-meal blood glucose (Arefeen et al., 3 Mar 2025). The reported cross-method comparison gives MealMeter a carbohydrate MAE of 13.2, RMSRE of 0.37, and correlation of 0.44; protein MAE of 9.66 and fat MAE of 3.67 (Arefeen et al., 3 Mar 2025). Here, the Feed-O-Meter notion becomes a nutrient-estimation engine driven by physiological response rather than direct audiovisual observation of food.
3. Closed-loop regulation and behavioral intervention
A major extension of the Feed-O-Meter concept is the transition from passive estimation to closed-loop control. “Earinter” exemplifies this shift by integrating in-the-wild sensing, real-time reasoning, and theory-grounded just-in-time intervention using commodity earbuds (Fang et al., 10 Feb 2026). The system repurposes the earbud’s bone-conduction voice sensor to capture chewing-related vibrations and operationalizes eating pace as chews per swallow, or CPS (Fang et al., 10 Feb 2026).
The sensing pipeline combines candidate segmentation over 50 ms frames with a 40 dB threshold, segment-length constraints of 100–400 ms, and a binary classifier inspired by EarVAS-Net. The classifier uses log-scaled Mel filter bank features of size 38 × 128 with EfficientNet-B0 as the feature extraction backbone, followed by mean pooling and a convolution, and is trained with focal loss (Fang et al., 10 Feb 2026). Chewing detection reaches Accuracy 96.86%, Precision 99.02%, Recall 95.05%, and F1 = 0.97 (Fang et al., 10 Feb 2026). Swallow detection is inferred heuristically from longer intervals between chewing bursts, and the system reports a macro-level chews-per-minute MAE of and a fine-grained CPS MAE of (Fang et al., 10 Feb 2026).
The closed-loop logic is theory-grounded in Dual Systems Theory. Before eating, the system delivers a System 2-oriented reminder; during eating, it provides short, System 1-compatible prompts; after eating, it presents a summary report (Fang et al., 10 Feb 2026). In a 13-day within-subject field study with , the mean CPS increased from 15.02 ± 3.67 at baseline to 26.42 ± 8.61 in Experiment, and food-consumption speed decreased significantly, with statistical signs of carryover on retention-probe days (Fang et al., 10 Feb 2026). This suggests a Feed-O-Meter can be not merely a meter but a regulator that senses, infers, intervenes, and re-senses within the same meal.
A related but distinct line is “MFED: A System for Monitoring Family Eating Dynamics,” which uses smart wearables, Bluetooth beacons, smartphones, a home base station, and EMA to monitor the context of eating in the wild in real time (Mondol et al., 2020). MFED does not identify what food is being eaten or how much is consumed, but it detects eating gestures and eating events, infers location and co-presence, and captures hunger, satiety, stress, mood, and social context through EMA (Mondol et al., 2020). The system was deployed in 20 homes with 74 participants, with responses from 4750 EMA surveys collected (Mondol et al., 2020). In this formulation, the meter tracks the ecology of eating rather than the nutritional contents or bite mass alone.
4. Feed-O-Meter as an educational feedback system
The explicit Feed-O-Meter paper defines the term in a pedagogical rather than dietary sense. The system is an LLM-based educational tool that helps design students practice giving feedback by role-playing as mentors to an AI mentee named Alex (Lim et al., 9 Sep 2025). The authors’ central concern is that many students struggle to provide feedback because of a lack of confidence, fear of being judged, and difficulty sustaining reflective back-and-forth critique in studio-based design courses (Lim et al., 9 Sep 2025).
The system is implemented as a web app with a React front end, Flask backend, GPT API integration, and MySQL logging, using GPT-4o for the main LLM modules with temperature set to 0 (Lim et al., 9 Sep 2025). Alex is described as a Korean first-year design major with limited design knowledge but a strong desire for feedback, and the knowledge state is explicitly constrained to grow only through user feedback (Lim et al., 9 Sep 2025). The pipeline includes a Categorizer, Knowledge extractor, Response generator, and Idea updater, and the interface includes onboarding, a main chat interface, and a Feedback Reflection Interface that displays facial expressions, inner thoughts, a knowledge level bar, and a Feedback Analysis Dashboard (Lim et al., 9 Sep 2025).
The feedback typology distinguishes six categories: Low-level question, Deep reasoning question, Generative design question, Share information, Evaluation, and Recommendation (Lim et al., 9 Sep 2025). For pipeline evaluation, 1,386 feedback sentences were collected; 60 were sampled for validation with 10 from each category; human-human agreement reached Cohen’s , and LLM vs. human agreement was and 0.72 against the two authors’ labels (Lim et al., 9 Sep 2025). In the within-subject user study with , Feed-O-Meter produced significantly longer feedback messages, with average syllables per user feedback of 207.85 versus 183.11 in baseline and (Lim et al., 9 Sep 2025). Expert evaluation found significantly higher scores for statement-based feedback on Specificity, Justification, and Action (Lim et al., 9 Sep 2025).
This usage departs semantically from food and animal feeding systems, yet the underlying systems pattern remains similar: an incoming stream is evaluated, transformed into interpretable quality indicators, and returned to the user as guidance for adaptive action.
5. Aquaculture and livestock interpretations
Outside human eating and design pedagogy, the Feed-O-Meter concept extends to feeding optimization in animal production. “Precision Aquaculture: An Integrated Computer Vision and IoT Approach for Optimized Tilapia Feeding” combines water-quality sensing, computer vision, IoT connectivity, and remote actuation to determine feed amounts for Tilapia (Hossam et al., 2024). The method estimates fish length using a YOLOv8 keypoint model and GLPN depth estimation, converts length to weight via
with and 0, maps weight to a daily feeding percentage by weight class, and multiplies by the counted number of fish (Hossam et al., 2024). The system also uses a YOLOv8 counting model, two synchronized cameras, an STM32F103C8 MCU, MQTT communication, a backend server, a mobile app, and a feeding pump mechanism with two servo motors and a 10 kg load cell with HX711 amplifier (Hossam et al., 2024). Reported performance includes keypoint Precision 94.50%, Recall 89.71%, AP@50 99.68%, AP@75 94.16%, counting Precision 96.21%, Recall 86.82%, AP@50 98.88%, AP@75 92.47%, and counting accuracy 94.5% on 100 frame pairs from actual farm conditions (Hossam et al., 2024).
“Towards Precision Feeding Using Behavioral Monitoring in Marine Cages” adopts a different strategy: rather than directly computing feed from fish morphometry, it infers feeding state from behavior (Georgopoulou et al., 2024). The system uses underwater video, YOLOv5 for fish detection, DEEPSORT for tracking, and group-level analysis to derive fish speed and a new Feeding Behavior Index (FBI) (Georgopoulou et al., 2024). Four feeding scenarios are studied: normal feeding, reduced feeding at 50% of normal, overfeeding at 150% of normal, and no feeding (Georgopoulou et al., 2024). Mean speed is reported as 0.54 ± 0.11 bd/sec under normal feeding, 0.61 ± 0.10 bd/sec under reduced feeding, 0.57 ± 0.12 bd/sec under overfeeding, and 0.56 ± 0.15 bd/sec under no feeding, with reduced feeding significantly higher than normal feeding at statistic = 4.48 and 1 (Georgopoulou et al., 2024). FBI is clustered with a Gaussian Mixture Model into four phases: Pre-feeding, Feeding start, Feeding later, and Post-feeding (Georgopoulou et al., 2024). This suggests a Feed-O-Meter can serve as a satiation predictor or feed-duration controller in sea cages.
A broader animal-monitoring perspective is provided by “Livestock feeding behaviour: A review on automated systems for ruminant monitoring” (Chelotti et al., 2023). That review defines feeding behavior hierarchically, from elemental events such as bites, grazing chews, rumination chews, and chew-bites, through eating and ruminating bouts, to daily grazing time, rumination time, intake rate, and dry matter intake (Chelotti et al., 2023). It surveys movement, acoustic, image/video, and pressure sensing systems, and emphasizes that no single sensing technology is universally best (Chelotti et al., 2023). The review explicitly frames a robust Feed-O-Meter for ruminants as a sensor-plus-algorithm system that must detect jaw movements reliably, distinguish bite, chew, chew-bite, and rumination chew, segment bouts, estimate feeding time, rumination time, intake rate, and ideally DMI, while remaining feasible on low-power embedded hardware (Chelotti et al., 2023).
6. Architectures, metrics, and recurring technical patterns
Despite the diversity of domains, several technical regularities recur across Feed-O-Meter-style systems.
| System | Primary signal | Core output |
|---|---|---|
| (Papapanagiotou et al., 2021) | Earbud audio + chew timestamps | Bite weight |
| (Kyritsis et al., 2020) | Smartwatch IMU | Bite timestamps and meal intervals |
| (Fang et al., 10 Feb 2026) | Earbud bone-conduction sensor | CPS and JIT intervention triggers |
| (Wang et al., 7 May 2025) | Egocentric RGB video + depth | Food volume |
| (Arefeen et al., 3 Mar 2025) | CGM + wrist physiological signals | Macronutrient quantities |
| (Hossam et al., 2024) | Vision + IoT sensors | Feed amount for Tilapia |
| (Georgopoulou et al., 2024) | Underwater video | Feeding state and satiation proxies |
| (Lim et al., 9 Sep 2025) | Multi-turn text interaction | Feedback-quality reflection |
One common pattern is multistage inference. In the earbud bite-weight system, the pipeline proceeds from chew timestamps and audio features to bout-level representations via Bag-of-Words or VLAD, then to regression using linear regression, SVR, FFNN, or GRNN (Papapanagiotou et al., 2021). In the smartwatch meal-monitoring system, the sequence is raw IMU, normalization and filtering, end-to-end bite detection, then signal processing on bite density to infer meal intervals (Kyritsis et al., 2020). In FoodTrack, the sequence is RGB capture, super-resolution, segmentation, temporal depth estimation, 3D reconstruction, metric scaling, then volume computation (Wang et al., 7 May 2025). In Feed-O-Meter for design education, the flow is user feedback, categorization, knowledge extraction, response generation, idea updating, and reflective presentation (Lim et al., 9 Sep 2025).
A second recurring pattern is the use of metrics that suit the estimated quantity. MAE and MAPE dominate bite-weight estimation (Papapanagiotou et al., 2021). F1 score and Jaccard Index dominate bite and meal localization (Kyritsis et al., 2020). Absolute percentage error is used for egocentric food-volume reconstruction (Wang et al., 7 May 2025). MAE, RMSRE, and Pearson correlation are used for macronutrient regression (Arefeen et al., 3 Mar 2025). Precision, Recall, AP@50, and AP@75 are used in aquaculture detection pipelines (Hossam et al., 2024). In educational Feed-O-Meter, the evaluation combines Cohen’s 2, expert Likert ratings, interaction logs, surveys, and interviews (Lim et al., 9 Sep 2025).
A third pattern is the tension between specificity and generalization. Food-specific bite-weight models outperform pooled models, with GRNN and FFNN achieving the strongest results when food type is known (Papapanagiotou et al., 2021). Smartwatch meal localization generalizes to external data, but performance drops on ACE because of differences in sampling rate, sensor setup, and behavior assumptions (Kyritsis et al., 2020). FoodTrack’s reported result is based on one sandwich example, so broader validation remains necessary (Wang et al., 7 May 2025). MealMeter is trained on only 12 healthy adults in controlled laboratory sessions (Arefeen et al., 3 Mar 2025). Aquaculture systems often depend on species-specific coefficients or feeding tables, such as 3 for Tilapia (Hossam et al., 2024).
7. Limitations, ambiguities, and future directions
The literature does not support a single, universally accepted definition of “Feed-O-Meter.” Instead, the term spans at least three senses: a named LLM-based feedback-training system (Lim et al., 9 Sep 2025), a family of food-intake and eating-behavior sensing systems (Papapanagiotou et al., 2021, Kyritsis et al., 2020, Fang et al., 10 Feb 2026, Wang et al., 7 May 2025, Arefeen et al., 3 Mar 2025, Mondol et al., 2020), and feeding-control systems in aquaculture and livestock monitoring (Hossam et al., 2024, Georgopoulou et al., 2024, Chelotti et al., 2023). A plausible implication is that the term functions more as a conceptual archetype than as a stabilized technical category.
Across the sensing literature, limitations are consistent. Manual annotation remains a bottleneck in bite-weight estimation, where chew boundaries and food type labels are required for training and evaluation (Papapanagiotou et al., 2021). Free-living meal localization still faces uncertainty because self-reported meal boundaries can be imprecise and some eating styles are underrepresented in training (Kyritsis et al., 2020). Earinter’s 13-day study is sufficient to demonstrate feasibility and short-term carryover, but not long-term habit change, and the use of a rough 25-chew heuristic is acknowledged as a simplification (Fang et al., 10 Feb 2026). FoodTrack is preliminary, with validation on one sandwich example and dependence on monocular depth estimation accuracy (Wang et al., 7 May 2025). MealMeter’s dataset is small and controlled, and protein estimation remains difficult (Arefeen et al., 3 Mar 2025). MFED does not estimate food identity or intake amount (Mondol et al., 2020). Aquaculture systems still do not fully integrate environmental factors into the feed computation itself or require further refinement for robust commercial deployment (Hossam et al., 2024, Georgopoulou et al., 2024).
In the design-education interpretation, the main limitations include reliance on multiple LLM modules, possible hallucinations, text-only interaction, small sample size, and cultural narrowness because all participants were Korean (Lim et al., 9 Sep 2025). The paper also reports a tension between creating a realistic, socially sensitive interaction and encouraging harder, more critical questioning (Lim et al., 9 Sep 2025).
Taken together, the current literature suggests that future Feed-O-Meter systems will likely be more multimodal, more context-aware, and more closed-loop. For human eating, this could mean integrating chewing detection, food-type recognition, bite-weight estimation, meal segmentation, and postprandial physiological inference into a unified passive monitoring stack. For aquaculture and livestock, it suggests tighter coupling between behavioral sensing, environmental monitoring, and real-time actuation. For education, it suggests more realistic agent modeling, multimodal critique interfaces, and longitudinal assessment of skill acquisition. The common research trajectory is from isolated measurement toward adaptive systems that not only estimate a state, but also support decisions or interventions based on that estimate.