KAP: Kinetic Augmented Pill Bottle for Vibration-Based Medication Interaction Recognition

This paper introduces KAP, an in-home medication interaction recognition system targeting unintentional medication non-adherence. KAP leverages kinetic-augmented pill bottles that have unique IDs embedded into their acoustic signatures for vibrations generated when they are put down. This enhanced pill bottle can be seamlessly integrated with existing vibration-based in-home occupant monitoring systems to detect and recognize medication interactions. Compared to prior work on medication non-adherence monitoring, such as using specially designed pill bottles, pill dispensers, or scanners, our system is battery-free and non-intrusive, while preserving the original usage of the standardized pill bottles from pharmacies. We design and implement KAP with a 3D-printed kinetic-augmentation attachment and a kinetic model-guided feature extraction and recognition algorithm. We conduct real-world data collection and depict preliminary results to demonstrate the feasibility of KAP.


INTRODUCTION
Physicians in modern healthcare prescribe medicine as a common tool to manage symptoms and diseases.Patients, however, vary widely on how strictly they follow their prescriptions [7].Because

Vibration Sensor
Pill bo/le interac1on But which bottle? medication adherence is an uncontrolled variable in modern-day medicine [18] and mostly relies on self-report assessment, it creates a real problem for not just the patients, but also physicians and caregivers.This is particularly a concern for older adults with cognitive impairment -a study of an elderly group with at least one medication shows a 17.4% of medication non-adherence rate [27].
Caregivers, who are in short supply in countries like America [25], end up with more burden to tend to patients' medication adherence.Meanwhile, physicians have difficulty managing patient medication plans and pinpointing the cause of medical emergencies, due to a potential lack of medication adherence [28].
The medication non-adherence problem stems mostly from unintentional rather than malicious non-compliance [14].One major reason that patients unintentionally deviate from their prescription plans is that they forget, with a 49.6% reporting forgetfulness as the major reason [11].
To mitigate the lack of medication intake monitoring, various Internet of Things systems have been developed.Customized onbottle sensing systems [2,19] often face battery issues, which are of high maintenance.Vision-based pill dispensers [6,22], on the other hand, would require the initial installation and constant refill of the dispenser.Furthermore, the RFID-and NFC-based near-field sensing approaches often require a designated platform/reader with specific usage requirements [15,17].Therefore, a battery-free and non-intrusive system that preserves the usage of the pill bottle are essential for medication interaction monitoring.
In contrast to these sensor approaches, surface kinetic-based sensing for human information inference has been well-developed in the last decade [8,20].The intuition is that when people interact with ambient surfaces, this contact will induce vibration.This vibration propagates through the surface and can be captured by vibration sensors nearby as shown in Figure 1.However, such indirect sensing often can extract limited information resolution.For example, it can be used to detect and recognize items picked up and put down when the item has different textures [31,33].However, when it is used for medication interaction monitoring, it may not be able to recognize what types of medication the patient interacts with due to the standardized pill bottle used at the pharmacy.
We present KAP, a kinetic-augmented pill bottle that can seamlessly work with existing surface vibration-based human sensing systems to track medication adherence through surface vibrations induced by setting down different pill bottles.By adding a lowcost identity base -the kinetic augmentation attachment -KAP requires no other retrofit to existing pill bottles and vibration-based activity recognition systems.The attachment embeds the identity information into the vibration signal generated by a human-bottle interaction (put down), allowing the augmented pill bottle to be battery free and low maintenance.A feature extraction algorithm based on the kinetic models is presented to achieve robust recognition of this embedded identity information.We summarize the contribution of this work as follows: • We propose to change the physics properties of ambient objects for information embedding in their kinetic signatures.• We design a kinetic-augmented pill bottle and corresponding recognition algorithm for unintentional medication nonadherence monitoring.• We conduct real-world data collection with 3D-printed kineticaugmentation attachments and present preliminary results to demonstrate feasibility.

BACKGROUND
In this section, we first introduce the motivation scenario and the scope of KAP.Then we depict the simplified physics model that describes the enabler of KAP.Finally, we present the result of a feasibility study.

Intentional vs Unintentional Medication Non-adherence
For our study, we define intention as the determination to act a certain way [16] while unintentional implies that the behavior is less associated with conscious decision-making and beliefs.
In our investigation for monitoring patient medication intake, we are not distinguishing between intentional and unintentional medication non-adherence.However, as a limitation of KAP, we understand that patients have to comply with this system in order to function properly.
However, particularly among long-term pharmacotherapy patients, KAP may have significant impacts on patient medicinetaking habits.Studies indicate that patients tend to exhibit improved medication adherence when provided with reminders or encouraged to form habits [21,24].

Ideal Bouncing Ball Model
In an ideal system, a perfectly elastic ball dropping and hitting an immovable rigid surface would deform it, where all the kinetic energy transforms to the elastic potential energy [26].Because of the conversation of momentum, if the surface does not move, the ball would bounce back [26], and in this process, the elastic potential energy converts back into kinetic energy.In the simplified ideal bouncing ball model, the coefficient of restitution (COR) is used to describe the elasticity of colliding objects in a system [10].COR  is described as the ratio of the final and initial relative speed between two objects  = The COR is determined by the material property Poisson's ratio  and the velocity at the impact , described as  ∝ 1/4 [10].The bouncing interval   and height  , as marked in Figure 2, are functions of , which demonstrates significant variance across different  configurations.Based on this observation, we design both the kinetic-augmented attachment to generate a signal with the corresponding features for more robust recognition.

Feasibility Study
To verify the feasibility of embedding a unique ID into the object's kinetic signature, we conduct a feasibility study illustrated in Figure 3(a).A string is attached to a fixed end and a bouncing ball.We lift the ball to a designated height and let it free fall.Then we capture the vibration with a geophone sensor placed on the same surface.Figure 3(b) and (c) depict the vibration generated by this ball drop when a metal ball and a rubber ball are used respectively.The blue lines mark the vibration signal captured, and the red dashed lines mark each bounce-induced impulsive signal.Each of these impulsive signals is generated when the height in Figure 2 is 0. The red solid line marks the general decay trend of these bounce-induced impulsive signals.We can observe that the bouncing interval value and variance, as well as the decay trend, vary over different ball bounces.This result verifies our idea to embed the identity into the object's vibration signatures by configuring the kinetic properties.

SYSTEM DESIGN
We envision KAP being used seamlessly with the existing vibrationbased occupant monitoring system [8,9,20].When the user interacts with the pill bottle by picking it up and putting it down, this interaction will induce the surface (e.g., nightstand, countertop) to vibrate [34].The vibration-based occupant monitoring system would capture this vibration for further inference on the occupant activities.KAP consists of two major modules -a signal augmentation attachment, and a vibration-based pill bottle recognition algorithm -as shown in Figure 4. First, KAP leverages a pill bottle attachment with a bouncing mass connected to it via a spring to generate unique bouncing/oscillation patterns.In this way, the medication ID is embedded in this signal augmentation attachment before it propagates through the surface and is captured by the vibration sensor (Section 3.1).Then, the pill bottle recognition module will extract features and establish a classifier for pill bottle identification.

Kinetic Augmentation Attachment (KAA) for Standard Pill Bottle
The goal of this attachment is to embed unique IDs into the signal generated by the interaction of the pill bottle when it is put down on a surface.We propose to achieve that by adding a bouncing attachment with designated kinetic properties that can provide additional observation when an event (here we focus on the putdown motion of the bottle) occurs.As shown in Figure 2, when the material properties of the bouncing ball are different, the bouncing time interval (  ) and bouncing height ( ) vary.Via different attachments, the user generates a periodic signal that is uniquely augmented through a mass attached to a spring.The vibrations of the augmentation follow a decaying energy formula, where the height of the mass decays by the squared restitution  2 embedded in the bouncing ball.Depending on different configurations of the attachment (e.g. the material of the colliding body), the frequency of the periodic impulses, the time between each impulse, as well as the kinetic vibration decay of each impulse will all depend on a unique  restitution value.This kinetic vibration decay propagated through the contacted surface is read by a geophone and collected for data processing.

Kinetic-Aware Feature Extraction (KAFE) and Pill Bottle Recognition
We obtain the pill bottle put down events detected from existing vibration-based activity recognition algorithm [20].We design a kinetic-aware feature extraction algorithm to effectively capture the kinetic characteristics of the augmentation structure.features that reflect the decay models of the corresponding KAA.
The parameters, such as the measurement of   and the clamped number of peaks detected for different thresholds (a function of  ), are fed into an RBF-kernel Support Vector Machine to recognize the particular augmentation attachment that generates the signal of the event.The detailed implementation is open-sourced at https: //github.com/dlee267/Hotmobile2024-KAPill.

EVALUATION
We conduct real-world experiments to evaluate KAP.

System Implementation
We 3D print our augmentation attachment as shown in Figure 4 with  1 = 43mm,  = 20mm,  = 20mm, and install different oscillation generators to fabricate different KAA configurations.We consider three configurations: (1) no augmentation, (2) a metal ball,  2 = 8mm, and (3) a rubber ball,  2 = 27mm.We adopt an RBF-kernel Support Vector Machine as a classifier to recognize pill bottles based on their vibration signatures.

Experiment Setup and Data Collection
Figure 6(a) shows the experiment setup, where five locations are selected on a wooden table of size 120cm × 60cm.The vibration sensor is vertically installed on the top-left corner of the table.Each pill bottle is put down in each location 50 times.Figure 6 (b-f) lists the investigated pill bottles.To verify the robustness of KAP on location variance, we select one location's data as testing data, and another four locations' data as training data (Leave One Out, LOO).

Baseline and Evaluation Metric
We compare KAP to two baselines: three pill bottles of different sizes without kinetic-augmentation (No KAA No KAFE); and without kinetic-aware feature extraction (No KAFE).We take prediction accuracy as the metric to evaluate the performance of pill bottle recognition.

Results and Analysis
Figure 7 shows the recognition accuracy of KAP (with KAA and KAFE) and two baselines (No KAFE, No KAA and No KAFE) at each testing location.We repeat the training and testing task 20 times and report the mean and the standard deviation of the learning accuracy.Our KAP (No LOO, 90% training data) achieves up to 99% (± 2.38%), 97% (± 3.94%), 98% (±2.89%), 100% (± 0.00%), 99% (± 2.00%) at the five testing locations respectively.KAP also demonstrates robustness against location variance in the LOO test.The accuracy of KAP (LOO) for the four locations M, T-R, B-R, and T-L achieves an average of 82% (± 5.64%), which is higher than No KAA and/or No KAFE, with a 20%, 27% average accuracy improvement, respectively.We observe that location B-L demonstrates a different trend, which might be caused by the structural abnormality that creates a significant heterogeneous area, which we will look into the causes in the future.

RELATED WORK
We compare KAP to relevant prior work.

Medication Intake Monitoring.
To monitor medicine intake, a wide variety of medication intake monitoring systems have been developed.First, customized pill bottle systems leverage on-bottle sensors, such as IMU [3] and magnetic sensors [2,19], to detect the interaction.The pill bottle identity is intrinsically included by the embedded system; however, the smart pill bottle requires batteries to power, therefore requiring constant maintenance.Second, vision-based pill dispensers are another common approach taken [6,22].However, the initial installation and the refill of the medication take additional effort, and it needs to be combined with other sensor components for accurate verification in different situations, such as patients wearing different colored clothing, or a bottle placed at a specific location [4].Third, wearable-based approaches often rely on sensor components worn on patients' necks or wrists to track the motion of swallowing [13], pill bottle opening, and pill removal [12], which is not practical for the elderly population, due to discomfort and regularly requirement of charging.Last, RFID-and NFC-based approaches have been explored [5,15,17], which do not require significant changes to the pill bottle other than a tag.For example, the SmartDrawer [5] uses a surface-mounted antenna to sense the proximity of RFID tags on pill bottles.These RFID-based systems often require an RFID reader to cover a designated area for medication intake monitoring tasks only.The KAP system, on the other hand, utilizes existing surface vibration-based monitoring systems.Therefore, the vibration-based system can detect pill-bottle interaction on any rigid surfaces at home.

Surface Sensing for Human Information
Monitoring.
Many sensing modalities or sensors have been designed and developed to capture human-surface interaction, such as FlexiForce Sensors [23], velostat sensors [30], piezoelectric sensors [29], laserbased vibrometery [32], and geophones [1].These sensors measure the deformation of the surface caused by human/object interaction and convert it into voltage via different principles.They are often used for human-surface contact detection and recognition.However, their recognition ability is limited based on their sensing principle.To enhance their sensing ability, instead of enhancing the sensor's resolution, we enhance the sensing target's physical properties to embed additional information into the target signal and achieve finer-grained monitoring.

DISCUSSION AND FUTURE WORK
In this section, we discuss the limitations and challenges of current implementations, as well as our future work plan.
Optimal Kinetic Augmentation Configuration.The form factor of our kinetic augmentation attachment for pill bottles is a 3D-printed base connecting to the mass ball via a spring.The structural properties (elasticity, size/length, shape) of the base, the spring, and the ball can be configured.Here, we investigate two types of balls with different sizes and materials as a feasibility check for physical information embedding.In the future, we plan to explore more configuration combinations and other forms of physical augmentations.For example, non-spring mechanisms similar to membrane switches may be incorporated in KAA for generating unique kinetic signatures.As a requirement for future configurations, it is be important to select a globally optimal configuration to maximize the number of pill bottles that can be identified with the constraint of available configurable parameter options.
Medication Amount Inference.We held an assumption that the interaction (put-down event) of the pill bottle indicates medication intake.If the person (especially dementia patients) picked it up and put it down without taking it, the KAP system would falsely detect a medication intake event.Our design scope is to provide additional aid to track medication for patients who may easily forget to take their medication, such as Alzheimer's disease, but are otherwise willing to take their medication.To provide higher-resolution tracking, we plan to further explore solutions to distinguish signal differences when the amount of medication in the bottle changes.For example, multitask learning may jointly learn the fill levels as well as the KAA-encoded medication types.
Robustness to Human Behavior Variance.In this work, we focus on the modeling of the KAA-surface contact and conduct experiments with one user.This controlled setting may not capture the human behavior variance inter-and intra-personally.For example, the same person may interact with the pill bottle differently when they are under different moods or health conditions over time, which generates intra-personal data shift.Different people may have different habits when putting down things, which causes inter-personal data shift.Therefore, one direction we plan to explore is to conduct long-term data collection (e.g.daily for over a month) with a larger number of users to quantify the inter-and intra-personal data distribution shift.Algorithms will be developed to preserve the kinetic-augmentation characteristics while minimizing the impact of the human variance.
Robustness to Ambient Vibration Sources.Other common household activities, such as interaction with keys, pens, and books, may read as false positives for medication intake.In addition, for events that occur during the same period, the signals captured by the vibration sensor may overlap.Since KAP uses KAA to generate a unique acoustic signature, we will explore leveraging the prior known physics model (e.g., in the form of signal templates) to effectively guide the signal separation and recognition.
Robustness for Different Surfaces.We mean for the KAP system to be configured to work on many different surfaces.With differing materials, shapes, and thicknesses of surfaces, vibrations may propagate differently.In this paper, we controlled the surface properties for experiments.However, in the future, we will explore algorithms that are surface-agnostic or adaptive based on the surface properties that can be learned from the vibration waveform.
Robustness to Mechanical Wear.The current KAP system uses mechanical metal springs attached to bouncing masses, which may experience metal fatigue and elastic wear over long-term usage.For future work, we will take into account these factors in both system characterization and design.For example, different springs and masses may result in different mechanical wear rates and therefore impact the system's lifetime, this can be considered as an additional constraint in the optimization process.

CONCLUSION
In conclusion, KAP introduces a kinetic-augmented pill bottle that works seamlessly with surface vibration-based human sensing systems for remote medication intake monitoring.Augmenting the user-generated vibrations allows our system to identify different medications across different locations.KAP uses this additional information to achieve the same monitoring as previous systems without required maintenance, complicated network installation, or changing patient behavior.We investigate different locations and different augmentations with real-world experiments.KAP achieves up to 99% accuracy in supervised learning prediction and an average of 82% accuracy in the LOO test, demonstrating up to a 27% improvement compared to baselines.

Figure 1 :
Figure 1: KAP motivation.Existing vibration-based monitoring systems can detect pill bottle interaction but cannot distinguish the pill bottles interacted with.

Figure 2 :
Figure 2: Simulation of the ideal bouncing ball's height changes over time with different restitution .The key differences are bouncing interval   and the height amplitude  , for each bounce.

Figure 3 :Figure 4 :
Figure 3: Feasibility study.(a) illustrates the feasibility test procedure.(b) the vibration signal induced by a constraint metal ball bouncing.(c) the vibration signal induced by a constraint rubber ball bouncing.The red dashed line marks the individual bounces, and the red solid line marks the decay of the bouncing signal amplitudes.
90% training (ours) KAP with LOO (ours) KAP no KAFE with LOO no KAA no KAFE with LOO

Figure 7 :
Figure 7: The recognition accuracy of KAP and two baselines over different locations on the table.The dark green bars depict KAP's performance when 90% data is used for training.The light green bars show accuracy with the Leave-One-Out (LOO) setting, where the model is trained with data from locations other than the testing location.The blue and red bars demonstrate the performance of baselines -without KAFE and without KAA -with LOO setting.