M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

IoT-Enabled Deep

Reinforcement Learning

for Adaptive Waste

Management in Hospital

Environments

ARTICLE HISTORY

Received 24 December 2025

Accepted 23 June 2026

Published 7 July 2026

Muhammad Masood Ul Rahman Usmani

Department of Computer Science

COMSATS University Islamabad, Sahiwal Campus

Sahiwal, Pakistan

muhammadmasoodulrahmanusmani@gmail.com

ORCID: 0009-0001-9948-111X

Rimsha Raﬁq

Department of Computer Science

COMSATS University Islamabad, Sahiwal Campus

Sahiwal, Pakistan

rimshach026@gmail.com

ORCID: 0009-0009-4002-0798

Makki Riaz Khan

Department of Information Technology

Bahauddin Zakariya University, Lodhran Campus

Lodhran, Pakistan

makkiriazkhanwapda@gmail.com

ORCID: 0009-0009-0698-3857

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

This work is licensed under a Creative Commons

Attribution-NonCommercial-ShareAlike 4.0 International License.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026

IoT-Enabled Deep Reinforcement Learning for

Adaptive Waste Management in Hospital

Environments

Muhammad Masood Ul Rahman Usmani

COMSATS University Islamabad, Sahiwal Campus

Department of Computer Science

Sahiwal, Pakistan

muhammadmasoodulrahmanusmani@gmail.com

Rimsha Raﬁq

COMSATS University Islamabad, Sahiwal Campus

Department of Computer Science

Sahiwal, Pakistan

rimshach026@gmail.com

Makki Riaz Khan

Bahauddin Zakariya University, Lodhran Campus

Department of Information Technology

Lodhran, Pakistan

makkiriazkhanwapda@gmail.com

Abstract—Healthcare industry produces considerable amount

of biomedical waste, which needs efﬁcient and safe ways of

handling. The research paper puts forward a scheme of IoT-

enabled DRL for hospital waste management. Unlike previously

known schemes, the suggested scheme incorporates IoT-based

sensing capabilities with a Proximal Policy Optimization agent.

The performance of the framework is analyzed by developing

a custom simulation environment based on the OpenAI Gym

library, wherein the process of waste production is modeled as

stochastic. According to the experimental results, the suggested

scheme of PPO surpasses its competitors in all key criteria.

Statistical validation using ANOVA and t-tests conﬁrms that

the improvements are signiﬁcant. The ﬁndings highlight the

potential of IoT-DRL integration for intelligent, adaptive, and

efﬁcient hospital waste management systems.

Keywords—Biomedical Waste Management, Internet of

Things (IoT), Deep Reinforcement Learning (DRL), Healthcare

Systems, Intelligent Optimization, Hospital Safety, Sustainabil-

ity.

I. INTRODUCTION

The rising number of healthcare institutions has worsened

the situation regarding biomedical and hospital waste man-

agement. Inadequate methods of separating and disposing of

hospital waste can pose severe threats to the environment and

raise the risks of disease transmission to healthcare providers

and communities nearby [1], [2]. Conventional separation by

healthcare providers can be laborious, prone to mistakes, and

dangerous to healthcare workers [3], [4].

Recent developments in IoT and AI technologies have

opened avenues towards new paradigms in smart waste

management. Low-cost technologies in IoT (e.g., ultrasonic

level sensors, RFID tags, and environment sensors) enable

waste level detection at all times and provide real-time

input to central processing units [5]–[7]. The application of

intelligent systems to already existing infrastructure in IoT

makes it possible to manage waste proactively and decrease

the chances of overﬂow and pollution in sensitive sectors like

hospitals [7].

Machine learning algorithms, speciﬁcally deep learning

networks such as EfﬁcientNet, ResNeXt, MobileNet, and

YOLO, have achieved signiﬁcant success in classifying dif-

ferent types of healthcare waste accurately with a clas-

siﬁcation accuracy of more than 95% [1], [4], [8]. Such

systems bring the process of proper healthcare waste segre-

gation at par with the national healthcare waste management

guidelines, thus preventing the risks of exposure to harmful

healthcare waste [8], [22].

However, despite such advancements, the majority of ex-

isting systems are still either reactive or limited to a labscale

environment. For instance, even though IoT-enabled systems

can send notiﬁcations once the bin is full, they do not have

the predictive models required for optimal collection route

planning [7], [19]. Similarly, vision-based sorting models,

although data-intensive, may fail to effectively adapt to a

healthcare setting [9], [10]. Thus, a critical need emerges for

adaptive and smart systems to adapt to the ever-changing

hospital waste production dynamics.

Deep Reinforcement Learning (DRL) is also an emerg-

ing promising technology for adaptive decision-making in

challenging and uncertain environments. In contrast to static

ML models, DRL agents are capable of learning optimal

waste management techniques such as dynamic routing,

vehicle allocation, and prioritization of bins through trial and

error interactions in the environment [11], [12]. Multi-agent

DRL techniques are also further extended and employed for

simultaneous allocation of waste collection vehicles that aim

to reduce energy costs, minimize delays in waste collection,

and avoid the formation of dangerous wastes [11], [13].

In a hospital setting, the adoption of IoT-capable DRL

models will be beneﬁcial in making hospital waste manage-

ment activities more predictive and adaptive. Real-time data

inputs from sensors, along with models for reinforcement

learning, such as Proximal Policy Optimization (PPO) and

Deep Q-Networks (DQN), will be useful in making hospital

waste management more adaptive by optimizing routes, min-

imizing overﬂow levels, and maintaining biomedical waste

regulations [7], [11].

This study introduces an IoT-assisted Deep Reinforcement

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

Learning framework for adaptive hospital waste management.

The framework is founded on the recent advancements in the

classiﬁcation of healthcare waste in [8], proactive IoT assis-

tance in hospital waste management in [7], and multiagent

reinforcement learning with the goal of smart cities in [11].

The contributions of this paper are:

• An IoT-based sensor system capable of continuously

tracking waste generation, availability of bins, and other

risk factors in hospital environments.

• An accurate model for classiﬁcation of different types

of medical waste (hazardous, infectious, general, sharps,

etc.) using deep learning techniques on images and

sensor data.

• A DRL agent responsible for learning adaptive policies

regarding waste collection, bin services, routing, and

disinfection or storage considering safety and cost as-

pects.

• Evaluation on hospital data to assess improved classiﬁ-

cation accuracy, lower cost and risk, and ability to adapt

to variable amounts of waste.

The second part highlights related literature on IoT, ma-

chine learning, and RL on Waste Management. The third

section reviews the proposed IoT-enabled DRL framework.

The fourth part highlights the experiment setup. The ﬁfth

section presents results discussion and comparison. The sixth

part provides conclusions and future directions.

II. R

ELATED WORK

Hospital and biomedical waste management has been an

increasingly growing problem has attracted a lot of interest

from the scientiﬁc community lately. This is because of

the associated effects of infection control, sustainability, and

efﬁciency. Traditional methods involving manual segregation

and collection at pre-documented times have been found to

be increasingly ineffective in dealing with the dynamic and

dangerous nature of hospital waste [1] [2]. As such, there has

been considerable interest in the use of emerging technologies

such as IoT, ML, and DRL for smart and automated hospital

waste management models.

A. Deep Learning Applied to Healthcare Waste Classiﬁcation

A recent area that has shown great promise for automatic

classiﬁcation of healthcare waste is based on the principles of

deep learning. Various research articles that applied transfer

learning with ResNeXt, EfﬁcientNet, and MobileNet models

have demonstrated classiﬁcation achievements above 90% for

different categories of healthcare wastes [1] [2]. Speciﬁcally,

YOLO models have also demonstrated their efﬁciency with

respect to real-time processing, with YOLOv5 and YOLOv8

models achieving above 95% classiﬁcation accuracy for

different categories of wastes with lower inference times,

making these models suitable for hospital settings [3] [4] [8].

Nevertheless, these models also rely heavily upon large-scale

healthcare datasets, and their applicability for classiﬁcation

with hospital settings has shown some challenges [9] [10].

B. IoT-Enabled Smart Waste Systems

IoT-Smart garbage management systems in the IoT en-

vironment are also explored in terms of urban waste col-

lection points (WCC) and smart cities. The usage of low-

cost sensor nodes and cloud computing contributes to re-

altime monitoring of the ﬁll level, temperature, humidity,

and environmental aspects of garbage bins in the IoT-based

system [5]– [7]. ProWaste technology, which involves the

application of machine learning algorithms and IoT sensors,

demonstrates prediction rates of above 99% in forecasting

situations of overﬂow at WCCs. The conceptual structures of

smart garbage bins in the IoT environment are also explored,

incorporating the use of sensors, compression systems, and

solar-based circuits to maximize effectiveness and minimize

costs [7] [11] [20] [21].

C. Reinforcement Learning and Optimization

Routing optimization for waste collection is an important

applications area in the healthcare sector, considering the con-

sequences of inefﬁcient waste collection, which can result in

accelerated health risk escalation. Conventional methods for

optimization, also originating in the Vehicle Routing Prob-

lem, have been adapted for superior effectiveness, thereby

incorporating stochastic optimization techniques [12] [19].

In the more recent past, reinforcement learning has been

employed for effective dynamic optimization in uncertain

settings for waste routing, thereby facilitating adaptability

according to real-time waste generation levels [11] [13].

Furthermore, multi-agent deep reinforcement learning setups

ensure greater ﬂexibility, incorporating vehicle dispatch and

management for multiple zones in an upscale healthcare

facility, thereby increasing energy efﬁciency and overall [11].

D. Integration in Healthcare Contexts

However, their application to a hospital environment is still

very limited despite signiﬁcant advances in general smart

waste management. Biomedical waste is unique in hazardous

material classiﬁcation, strict regulatory compliance, and in-

fection risks for healthcare staff [8] [22]. IoT-enabled sensing

with DRL can help build adaptive, predictive frameworks that

manage the complexity of the waste system of a hospital.

In addition, such integration ﬁts into the sustainability and

public health goals while ensuring compliance with the

national and international biomedical waste standards [2], [7],

[11].

Summarily, from the literature, there are promising fronts

but also some identiﬁed gaps. The existing systems based

on IoT are mostly on notiﬁcation, and classiﬁcations based

on deep learning focused on correctness in lieu of optimized

routes. The application of reinforcement learning, although

a robust method, is less considered in a healthcare setting

scenario. There consequently arises a need for a uniﬁed

framework of DRL using IoT-enabled adaptive hospital waste

management.

Table I summarizes recent approaches in IoT, machine

learning, and DRL-based waste management systems. The

comparison shows the limitations of the existing methods,

motivating the need for an integrated IoT-DRL framework.

III. M

ETHODOLOGY

The proposed framework utilizes a combination of IoT

sensing and DRL optimization techniques to provide an adap-

tive biomedical waste management system within a hospital

setting. The methodology is comprised of three fundamental

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026

TABLE I. Comparative Summary of Recent IoT, ML/DL, and DRL Approaches for Waste Management (2021–2025)

Authors / Year Method / Model Application / Dataset Key Results / Findings

Zhou et al. (2022) [1] ResNeXt-50 (DL) Private 8-class medical waste

dataset

Achieved 97.2% accuracy in healthcare waste

classiﬁcation.

Kumar et al. (2021) [2] EfﬁcientNet-B7 (TL, DL) COVID-related biomedical waste

streams

Reported 99% accuracy; highlighted AI for

circular economy in healthcare waste.

Kunwar & Rai (2025)

[8]

YOLOv5-s, YOLOv8,

EfﬁcientNet-B0

Medical Waste Dataset 4.0

+ Pharma-biomedical dataset

(Nepal)

YOLOv5-s achieved 95.06% accuracy; de-

ployed with bin-color mapping to Nepal’s

HCW standards.

Mok (2024) [4] YOLO + IoT Integration 67,860 images of HCW detection Achieved mAP of 98%; demonstrated IoT-

enhanced automated sorting.

Lahoti et al. (2024) [3] Computer Vision + Robotic

Arm

Multi-class waste segregation

prototype

Enabled real-time robotic segregation of hos-

pital waste.

Stephan et al. (2025)

[6]

ProWaste (IoT + ML, Deci-

sion Tree + BPSO)

Urban WCCs, 6,954 daily

records (Bengaluru)

Reduced missed pickups; >99.8% macro-F1

with only 3 predictive features; deployed via

mobile app.

Patil et al. (2021) [5] Ultrasonic Sensors + IoT E-waste and bin monitoring pro-

totype

IoT-based notiﬁcation system reduced manual

inspections.

Shanthini et al. (2021)

[7]

IoT, RFID, WSN Smart City Waste Collection Found LoRaWAN-based IoT systems outper-

form others in efﬁciency.

Rajani et al. (2022)

[11]

Multi-Agent Deep RL

(DRL)

IoT-driven smart waste manage-

ment (simulation)

Proposed platform-agnostic DRL framework;

optimized routing, reduced overﬂow.

Mishra et al. (2022)

[12]

Route Optimization (VRP +

IoT)

Bhubaneswar City MSW data Reduced vehicle distance by 30.28%, OpEx

by 29.07%.

Abuga et al. (2022)

[13]

IoT + Fuzzy Logic Real-time smart garbage bins Achieved high reliability in dynamic bin mon-

itoring and waste-level prediction.

Gondal et al. (2023)

[9]

Hybrid Deep Learning

Model

Real-time garbage classiﬁcation Achieved 99% training/validation accuracy

with automated bin sorting.

Zhang et al. (2022)

[10]

Cascade R-CNN (enhanced

with dilated convolutions)

Garbage detection for small ob-

jects

Improved precision in detecting small waste

objects in cluttered environments.

components, namely: (i) IoT Layer, (ii) DRL Layer, and (iii)

System Workﬂow. Figure 1 shows the overall architecture.

Fig. 1: Proposed IoT-Enabled DRL Framework for Adaptive

Hospital Waste Management

A. IoT Layer

This layer contains a network of intelligent biomedical

waste dumpsters with ultrasonic sensors for measuring the

ﬁll levels of the dumpsters, an RFID for identifying biowaste

types, and temperature sensors for the detection of infec-

tious diseases. Data from the dumpers is relayed through

lightweight messaging transports such as MQTT to a cloud

server. This enables real-time mapping of biowaste generation

behavior in hospital wards, operation theatres, and labs.

Figure 2 illustrates the PPO architecture used in this study.

Fig. 2: Proposed PPO Model Architecture

B. DRL Layer

The DRL layer formulates waste management as a sequen-

tial decision-making problem. Every individual in charge of

collecting waste material from different hospital locations

(for instance, a robot cart or human personnel) is considered

an agent functioning within the hospital premises. The state

space will consist of bin levels, location, waste importance

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

Algorithm 1 Training Procedure of the Proposed PPO Agent

1: Randomly initialize the policy parameters θ and value

network parameters ϕ

2: for each training episode do

3: Reset the environment and obtain the initial state s

4: for each interaction step until horizon T do

5: Sample an action a according to the current policy

(·|s)

6: Apply a to the environment

7: Receive reward r and the subsequent state s

′

8: Save the transition (s, a, r, s

′

) in the rollout buffer

9: Set s ← s

′

10: end for

11: Estimate discounted returns and compute advantages

using GAE

12: Evaluate the clipped objective function:

PPO

= E



min



r(θ)

A, clip(r(θ), 1 − ϵ, 1+ϵ)



13: Optimize the policy network using gradient ascent

14: Update the value network by minimizing the value

prediction error

15: end for

(hazardous or non-hazardous), and availability of employees.

Actions will comprise routing actions (next bin to collect,

segregation of waste material, or inaction). Reward functions

will aim at (i) reducing the routing distance and duration of

waste collection; (ii) avoiding overﬂowing of bins containing

hazardous waste materials; and (iii) compliance with the

biomedical safety standards. PPO will be chosen as the

algorithm to use based on DRL since it is stable and suited

for continuous state spaces.

C. System Workﬂow

The overall system workﬂow is as follows:

• IoT sensors collect real-time data from hospital waste

bins.

• Data is transmitted to a cloud server and preprocessed

(normalization, bin classiﬁcation).

• The DRL agent receives the current hospital state and

selects an optimal action (e.g., which bin to service).

• The action is executed by the waste collection system

(robot/human).

• The environment returns feedback (updated state, reward

signal).

• The DRL model updates its policy iteratively through

training episodes.

This adaptive loop allows the system to dynamically opti-

mize collection routes and schedules while minimizing risks

of contamination and operational inefﬁciencies.

1) PPO Training Algorithm: Use Proximal Policy Opti-

mization (PPO) to train the DRL agent [14]. Algorithm 1

presents the pseudocode of the training process.

D. Simulation Environment

For assessing the proposed framework, a simulation envi-

ronment was designed with the help of the OpenAI Gym API.

The environment considers hospital waste bins, collectors,

Algorithm 2 Custom OpenAI Gym Environment for Hospital

Waste Management

Require: Number of bins B, Hospital layout L, Maximum

steps T

Ensure: State transitions, reward signals

1: Initialize: Bin ﬁll levels f

=0, Waste types w

∈

{hazardous, non-hazardous}, Agent position p

2: for Epoches =1to N do

3: Reset: f

∼ U(0, 0.3), p

= nurse station

4: for t =0to T − 1 do

5: State: s

= {f

}

6: Agent selects action a

∈

{move to bin, collect, idle}

7: if action == collect then

8: Empty selected bin, update f

← 0

9: Reward r

=+α (hazardous) or +β (non-

hazardous)

10: end if

11: Update ﬁll levels: f

← f

+ δ

12: if f

> 1.0 then

13: Penalize: r

= −γ (overﬂow)

14: end if

15: Update agent position p

t+1

and hospital state

16: end for

17: if all bins empty or t = T then

18: Terminate episode

19: end if

20: Return trajectory {s

t+1

} for PPO training

21: end for

and IoT sensors dynamics. In each episode, the simulation

reﬂects the dynamics of a hospital shift, where the agent tries

to ﬁnd an optimal path and waste segregation strategy. The

algorithm for the environment is presented in Algorithm 2.

I V. R

ESULTS &EXPERIMENTS

This section outlines the experimental conﬁguration, ex-

plains the evaluation criteria, and analyzes the performance

of IoT-based DRL framework for adaptive hospital waste

management system. The experiments have been performed

using the Python environment that uses the custom OpenAI

Gym environment described in Algorithm 2 and TensorFlow

for DRL. Experiments were conducted on a workstation with

Intel i7-12700 CPU, 32GB RAM, and NVIDIA RTX 3080

GPU.

A. Experimental Setup

The hospital environment was simulated with B = 50 bins

distributed across wards, laboratories, and operation theaters.

Each bin was randomly assigned as hazardous (30%) or

non-hazardous (70%), and waste generation rates followed

a Poisson distribution. It is important to note that the exper-

imental evaluation is conducted using a simulated hospital

environment. IoT data streams are synthetically generated to

emulate real-world waste generation patterns. The DRL agent

(PPO) was compared against baseline methods:

• Rule-Based Scheduling (RBS): Fixed-time collection

intervals.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026

TABLE II. Simulation and PPO Hyperparameters

Parameter Value

α (Hazardous reward) 10

β (Non-hazardous reward) 5

γ (Overﬂow penalty) 15

λ (Poisson rate) 0.2–0.5

Learning Rate 3 × 10

−4

Discount Factor (γ) 0.99

PPO Clipping (ϵ) 0.2

Batch Size 64

Epochs 10

TABLE III. Performance Comparison of Waste Management

Approaches

Method Overﬂow

(%)

Coll.

Time

(min)

Haz.

Score

Energy

(kWh)

RBS 18.2 12.5 0.62 4.8

SPH 12.9 9.4 0.71 4.1

DQN 9.7 8.1 0.76 3.7

Proposed PPO 5.4 6.9 0.89 3.2

• Shortest Path Heuristic (SPH): Greedy routing to the

nearest non-empty bin.

• Deep Q-Network (DQN): Model-free baseline with dis-

crete actions.

B. Hyperparameter Settings

To ensure reproducibility, the following hyperparameters

were used:

C. Metrics for Evaluation

The system was evaluated using the following metrics:

• Overﬂow Rate (%): Percentage of bins that exceeded

capacity.

• Average Collection Time (min): Mean time per bin

service.

• Hazardous Waste Priority Score: Ratio of hazardous

waste collected before overﬂow.

• Energy Efﬁciency (kWh): Energy consumed by collec-

tion robots/vehicles.

D. Results and Discussion

Table III shows the comparison across baseline methods.

The proposed PPO-based DRL outperformed all baselines,

achieving the lowest overﬂow rate and highest hazardous

waste priority compliance. Fig. 3 and Fig. 4 illustrate the

overﬂow reduction and training convergence, respectively.

The ﬁndings validate that DRL-based adaptive scheduling

helps minimize the risks of overﬂow as well as enhances

the efﬁciency of hazardous waste management. Additionally,

the reduced energy consumption indicates the sustainability

advantages of such an approach in practical settings of

hospitals as well.

E. Statistical Signiﬁcance Analysis

We further conducted statistical signiﬁcance tests to assess

the robustness of the outcomes of the proposed method

and the baselines. We executed 30 independent episodes for

each method based on randomly generated patterns of waste

production. A one-way ANOVA test is followed by pairwise

two-tailed t-tests using Bonferroni correction for overﬂow

rate.

Fig. 3: Overﬂow reduction across training episodes

Fig. 4: Training convergence of PPO agent

Table IV shows that the ANOVA test yielded a statistically

signiﬁcant difference among methods (p<0.001). Post-hoc

pairwise t-tests (Table V) conﬁrmed that the proposed PPO

signiﬁcantly outperformed RBS, SPH, and DQN.

TABLE IV. One-Way ANOVA Results for Overﬂow Rate

Across Methods

Source DF F-Value p-Value

Between Groups 3 24.56 < 0.001

Within Groups 116 – –

Total 119 – –

TABLE V. Pairwise t-Test Results (Overﬂow Rate, Bonfer-

roni Corrected)

Comparison t-Statistic p-Value

PPO vs. RBS 7.82 < 0.001

PPO vs. SPH 6.11 < 0.001

PPO vs. DQN 3.45 0.002

The statistical analysis conﬁrms that the improvements of

PPO are not due to random chance, but are signiﬁcant at

the 95% conﬁdence level. Thus, the proposed DRL-based

approach provides a robust and reliable optimization strategy

for hospital waste management.

F. Complexity Analysis

To assess the computational feasibility of the proposed

framework, we analyzed the time and space complexity of

both the IoT data acquisition and the DRL training phases.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

1) IoT Communication Overhead: The IoT layer relies

on MQTT-based communication for transmitting sensor data

from B bins. Each message has an average payload size

of O(1) (bin ﬁll level, type, timestamp). Thus, the per-step

communication complexity is

O(B) (1)

For a hospital with B = 100 bins, the per-second com-

munication load remains within 50–100 KB, which is well

within the capabilities of low-cost WiFi/LoRa gateways.

2) PPO Training Complexity: The PPO algorithm involves

two key components: trajectory collection and policy updates.

Let N be the number of steps per episode, E the number of

episodes, and M the number of policy update iterations.

• Trajectory Collection: Each step requires state evalua-

tion and action selection, with complexity O(d), where d

is the dimensionality of the state vector. Thus, trajectory

collection over N steps per episode costs

O(Nd) (2)

• Policy Update: PPO uses stochastic gradient descent

on batches of size b over M iterations. Each forward-

backward pass has complexity O(θ) , where θ is the

number of neural network parameters. Thus, update

complexity is

O(Mbθ) (3)

Overall training complexity is therefore

O(END + EMbθ) (4)

3) Space Complexity: The memory footprint consists of:

replay buffer O(Nd), neural network parameters O(θ), and

IoT data queue O(B). Total space complexity is

O(Nd+ θ + B) (5)

In practice, with N = 1000 steps, E = 500 episodes,

M = 10 iterations, b = 64, and θ ≈ 10

, training can be

completed using a single NVIDIA RTX 3060 GPU in under 4

hours. The IoT communication overhead is negligible relative

to network capacity. Thus, the framework is computationally

efﬁcient and feasible for real-world hospital environments.

V. D

ISCUSSION

The obtained results conﬁrm that the proposed IoT-enabled

PPO framework provides a signiﬁcant improvement in hos-

pital waste management efﬁciency compared to baseline

approaches, including Rule-Based Scheduling (RBS), Short-

est Path Heuristic (SPH), and Deep Q-Network (DQN).

Speciﬁcally, the proposed model achieves lower overﬂow

rates, reduced collection times, and improved prioritization

of hazardous waste. These improvements are statistically

validated using ANOVA and post-hoc t-tests, conﬁrming

that the observed performance gains are not due to random

variation.

A. Interpretation of Results

These performance gains can be explained by two major

reasons. Firstly, the IoT-based data acquisition layer allows

for the adaptive tracking of waste production dynamics and

thus helps make the decisions in a timely fashion depending

on current conditions in the hospital, which is more efﬁcient

than static or heuristic methods.

Secondly, the PPO method offers stable learning in

stochastic environments thanks to the clipped surrogate ob-

jective function that does not allow for very large updates

of the policy function during learning. This feature is espe-

cially important in hospitals due to the uncertainty in waste

production and other constraints.

B. Limitations

However, despite all the positive results achieved, some

limitations should be mentioned. First of all, the evaluation of

the proposed solution takes place within a simulation setting

wherein the patterns of waste generation are artiﬁcially

generated. Even though these patterns are created with an

intention to mirror realistic conditions at the hospital, they

might not be sufﬁciently diverse enough.

Waste generation in real life may be subject to some unex-

pected factors like outbreaks of infectious diseases, changes

in seasons, and the increase in patient ﬂow, among others. In

particular, during a large-scale outbreak, the volume and con-

tent of biomedical waste may change substantially. Hence, the

evaluation of the system that takes place within a simulation

is an idealistic case study.

C. Ethical and Privacy Considerations

This approach uses information from hospital waste, some

of which might be conﬁdential metadata, like timestamps

that may inadvertently correspond to patient behavior. Com-

pliance with data protection laws, HIPAA and GDPR, is a

prerequisite before implementation of such an approach.

In addition, although automation decreases the exposure

to harmful waste for people, it is vital to retain proper

human control. System malfunctions, wrong classiﬁcations or

any other circumstances can be dangerous. Thus, the ethical

implementation of the system requires proper mechanisms of

monitoring and fail-safes.

D. Generalizability and Future Deployment

Though the model has been designed with hospital waste

management in mind, the suggested IoT-DRL model can be

easily customized for any other type of waste management

operations, such as intelligent cities and industries. The

implementation of multi-agent reinforcement learning will

help scale the model even more, as it will allow coordination

among several collectors.

Further research should be directed at applying the model

in practice through collaboration with the healthcare sector.

Besides, introducing uncertainty analysis, anomaly detection,

and edge computing is a good idea as well.

VI. C

ONCLUSION AND FUTURE WORK

This paper provided a novel IoT based deep reinforcement

learning (DRL) approach to adaptive hospital waste manage-

ment. The proposed method consists of using an IoT based

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July - December 2026

sensor along with proximal policy optimization (PPO) for

dynamic optimization of waste collection, route planning, and

prioritization of hazardous wastes.

From the results of our experiments performed in the

custom simulation environment built on top of the OpenAI

Gym, we can see that the behavior of the proposed DRL

method is signiﬁcantly better than those of the baselines

methods RBS, SPH, and DQN. Our model provides the

results with lower overﬂow rate, lesser collection time, and

better prioritization of hazardous wastes.

The primary contributions of this study are summarized as

follows:

• Development of an integrated IoT-DRL framework for

adaptive hospital waste management.

• Design of a custom simulation environment to model

realistic waste generation and collection dynamics.

• Comprehensive performance evaluation demonstrating

the effectiveness of PPO in complex and stochastic

environments.

• Analysis of computational complexity and consideration

of ethical and deployment-related aspects.

However, there are some limitations of this study. The

assessment was conducted based on artiﬁcial datasets in

a simulation setting, which does not completely account

for real-life randomness in hospital operations. Hence, the

obtained results correspond to an ideal case scenario.

The future works will be directed towards practical im-

plementation of the proposed approach together with hospi-

tals. Furthermore, the use of the multi-agent reinforcement

learning may make the system capable of collecting waste

cooperatively via several autonomous agents in hospital areas.

Some additional improvements might include robustness en-

hancement via modeling of uncertainties, anomalies detection

and failure-proofed IoT communications.

Additionally, the employment of energy-efﬁcient DRL

models and edge computing methods will enhance the scala-

bility and application of the framework in limited-resource

settings. Moreover, apart from being applied to hospitals,

the proposed framework can potentially be used for other

applications such as smart city waste management systems,

industrial waste handling systems and other public health

infrastructure.

Thus, the combination of IoT and DRL may lead to the

development of an advanced waste management framework.

UNDING STATEMENT

There was no outside support for this study.

CKNOWLEDGMENT

The Department of Computer Science at COMSATS Uni-

versity Islamabad’s Sahiwal Campus is acknowledged by the

authors for providing the academic resources and assistance

required to complete this study.

UTHOR CONTRIBUTIONS

The authors’ contributions follow the CRediT (Contributor

Roles Taxonomy) as follows:

• Muhammad Masood Ul Rahman Usmani: Conceptu-

alization, Methodology, Writing Original Draft & Edit-

ing

• Rimsha Raﬁq: Data Curation, Visualization, Review &

Editing

• Makki Riaz Khan: Conceptualization, Writing – Re-

view & Editing

EFERENCES

[1] H. Zhou, X. Yu, A. Alhaskawi, Y. Dong, Z. Wang, Q. Jin, X. Hu,

Z. Liu, V. G. Kota, M. H. Abdulla, S. H. A. Ezzi, B. Qi, J. Li, B.

Wang, J. Fang, and H. Lu, “A deep learning approach for medical

waste classiﬁcation,” Scientiﬁc Reports, vol. 12, no. 1, p. 2159, 2022.

[2] N. M. Kumar, M. A. Mohammed, K. H. Abdulkareem, R. Damase-

vicius, S. A. Mostafa, M. S. Maashi, and S. S. Chopra, “Artiﬁcial

intelligence-based solution for sorting COVID related medical waste

streams and supporting data-driven decisions for smart circular econ-

omy practice,” Process Safety and Environmental Protection, vol. 152,

pp. 482–494, 2021.

[3] J. Lahoti, J. Sn, M. V. Krishna, M. Prasad, R. Bs, N. Mysore, and J.

S. Nayak, “Multi-class waste segregation using computer vision and

robotic arm,” PeerJ Computer Science, vol. 10, p. e1957, 2024.

[4] M. H. Mok, “YOLO combined with IoT for detection of healthcare

waste,” Applied Sciences, vol. 14, no. 3, p. 1167, 2024.

[5] U. Patil et al., “IoT based smart waste management system,” in

Proc. IEEE Int. Conf. Inventive Research in Computing Applications

(ICIRCA), 2021, pp. 1–6.

[6] T. Stephan, S. M. Hari Krishna, C.-C. Lin, U. Sumesh, S. Agarwal,

and H. Kim, “ProWaste for proactive urban waste management using

IoT and machine learning,” Scientiﬁc Reports, vol. 15, no. 1, p. 27790,

2025.

[7] S. Vishnu, S. R. J. Ramson, S. Senith, T. Anagnostopoulos, A. M.

AbuMahfouz, X. Fan, S. Srinivasan, and A. A. Kirubaraj, “IoT-enabled

solid waste management in smart cities,” Smart Cities, vol. 4, no. 3,

pp. 1004–1017, 2021.

[8] S. Kunwar and P. Rai, “Healthcare waste classiﬁcation using deep

learning aligned with Nepal’s bin color guidelines,” arXiv preprint

arXiv:2508.07450, 2025.

[9] A. U. Gondal, M. I. Sadiq, T. Ali, M. Irfan, A. Shaf, M. Aamir, M.

Shoaib, A. Glowacz, R. Tadeusiewicz, and E. Kantoch, “Real time

multipurpose smart waste classiﬁcation model for efﬁcient recycling

in smart cities using multilayer convolutional neural network and

perceptron,” Sensors, vol. 21, no. 14, p. 4916, 2021.

[10] C. Zhang, X. Zhang, D. Tu, and Y. Wang, “Small object detection using

deep convolutional networks: Applied to garbage detection system,”

Journal of Electronic Imaging, vol. 30, no. 4, p. 043013, 2021.

[11] K. R. Rajani, A. Gaddam, and J. Gaddam, “IoT-based smart waste

management using deep reinforcement learning,” Preprints, 2022.

[12] A. Mishra, S. Ghosh, and D. P. Jena, “Internet of Things based waste

management system for smart cities: A real time route optimization for

waste collection vehicles,” International Journal of Scientiﬁc Research

in Computer Science and Engineering, vol. 6, no. 1, pp. 37–42, 2019.

[13] D. Abuga and N. S. Raghava, “Real-time smart garbage bin mechanism

for solid waste management in smart cities,” Sustainable Cities and

Society, vol. 75, p. 103347, 2021.

[14] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov,

“Proximal Policy Optimization Algorithms,” arXiv:1707.06347, 2017.

[15] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.

Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski,

S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran,

D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through

deep reinforcement learning,” Nature, vol. 518, pp. 529–533, 2015.

[16] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction,

2nd ed. MIT Press, 2018.

[17] Y. Li, “Deep reinforcement learning: An overview,” arXiv preprint

arXiv:1701.07274, 2017.

[18] L. D. Xu, W. He, and S. Li, “Internet of Things in industries: A survey,”

IEEE Transactions on Industrial Informatics, vol. 10, no. 4, pp. 2233–

2243, 2014.

[19] A. Martikkala, B. Mayanti, P. Helo, A. Lobov, and I. F. Ituarte,

“Smart textile waste collection system—Dynamic route optimization

with IoT,” Journal of Environmental Management, vol. 335, p. 117548,

2023.

[20] S. R. J. Ramson, D. J. Moni, S. Vishnu, T. Anagnostopoulos, A. A.

Kirubaraj, and X. Fan, “An IoT-based bin level monitoring system

for solid waste management,” Journal of Material Cycles and Waste

Management, vol. 23, no. 2, pp. 516–525, 2021.

[21] D. Baldo, A. Mecocci, S. Parrino, G. Peruzzi, and A. Pozzebon, “A

multi-layer LoRaWAN infrastructure for smart waste management,”

Sensors, vol. 21, no. 8, p. 2600, 2021.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

https://doi.org/10.33333/lajc.vol13n2.06

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

[22] S. Karki, S. R. Niraula, and S. Karki, “Perceived risk and associated

factors of healthcare waste in selected hospitals of Kathmandu, Nepal,”

PLOS ONE, vol. 15, no. 7, p. e0235982, 2020.

[23] H. Wu, F. Tao, and B. Yang, “Optimization of vehicle routing for waste

collection and transportation,” International Journal of Environmental

Research and Public Health, vol. 17, no. 14, p. 4963, 2020.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

AUTHORS

Muhammad Masood ul Rahman Usmani received his M.S. degree in

Computer Science from COMSATS University Islamabad, Sahiwal

Campus, Pakistan, in 2024, following his Master of Computer Science

degree from The Islamia University of Bahawalpur, Pakistan. He

is currently serving as a Faculty Member at Bahauddin Zakariya

University, Lodhran Sub Campus, Pakistan, where he teaches

undergraduate courses in Computer Science. His research interests

include Machine Learning, Deep Learning, Reinforcement Learning, the

Internet of Things (IoT), Computer Vision, and Artiﬁcial Intelligence for

healthcare. He has authored and coauthored several research articles

published in or submitted to international journals and conferences.

His current research focuses on intelligent IoT systems, medical image

analysis, and deep reinforcement learning based optimization for

smart environments.

Rimsha Raﬁq received her M.S. degree in Computer Science from

COMSATS University Islamabad, Sahiwal Campus, Pakistan, in 2024,

following her Bachelor degree in Computer Science from Bahauddin

Zakariya University, Multan, Pakistan. She is currently serving as a

Teaching Faculty at COMSATS University Islamabad, Sahiwal Campus,

Pakistan, where she teaches undergraduate courses in Computer

Science. Her research interests include Machine Learning, Deep

Learning, Reinforcement Learning and Artiﬁcial Intelligence for

healthcare. Her current research focuses on medical image analysis

and deep reinforcement learning based optimization for smart

environments.

Muhammad Masood Usmani

Rimsha Raﬁq

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 2, July 2026

AUTHORS

Makki Riaz Khan is a BS Information Technology student at Bahauddin

Zakariya University, Sub Campus Lodhran, Pakistan. His research

interests include Artiﬁcial Intelligence, Machine Learning, Deep

Learning, and Computer Vision. He is a co-author of the paper “Multi-

Class Classiﬁcation of Alzheimer's Impairment using EcientNet-B0”.

His technical expertise includes Python, TensorFlow, Keras, Scikit-

learn, and OpenCV. He has developed several AI-based applications

in healthcare and computer vision and is passionate about applying

intelligent technologies to solve real-world problems.

Makki Riaz Khan

M. M. U. R. Usmani, R. Raﬁq, and M. R. Khan,

“IoT-Enabled Deep Reinforcement Learning for Adaptive Waste Management in Hospital Environments”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 2, 2026.