6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

(http://www.acm.org)
☰ Article Navigation
(http://www.acm.org)

AI-Driven Healthcare Delivery in Pakistan: A

Framework for Systemic Improvement
Imama Zahoor (https://orcid.org/0009-0008-9058-3102), Lahore University of Management Sciences,
Pakistan, [email protected] (mailto:[email protected])
Shiza Ihtsham (https://orcid.org/0009-0004-7237-5597), Lahore University of Management Sciences,
Pakistan, [email protected] (mailto:[email protected])
Muhammad Umar Ramzan (https://orcid.org/0009-0004-8215-2928), Lahore University of
Management Sciences, Pakistan, [email protected] (mailto:[email protected])
Agha Ali Raza (https://orcid.org/0000-0003-0124-9783), Lahore University of Management Sciences,
Pakistan, [email protected] (mailto:[email protected])
Basmaa Ali (https://orcid.org/0000-0003-4900-1920), Lahore University of Management Sciences,
Pakistan, [email protected] (mailto:[email protected])

DOI: https://doi.org/10.1145/3674829.3675058 (https://doi.org/10.1145/3674829.3675058)

COMPASS '24: ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies
(https://doi.org/10.1145/3674829), New Delhi, India, July 2024

In Low- and Middle-Income Countries (LMICs), poor health outcomes come from a high burden of disease, a shortage of healthcare
professionals, and inefficient health information exchange leading to substantial economic losses. In this paper, we highlight critical gaps in
healthcare delivery in Pakistan and propose solutions to improve patient outcomes in resource-constrained environments. We have built
Darcheeni, an AI-driven healthcare framework that leverages artificial intelligence to assist and supplement physicians, streamline healthcare
processes, and prioritize patient-centered care. Darcheeni analyzes doctor-patient interactions in real-time, integrates lab and imaging data,
generates and distributes care plans customized to the patient's needs, and sends them directly to patients’ smartphones. We also discuss the
challenges and limitations associated with sustainable AI integration by centering our learnings from the pilot deployment of Darcheeni. By
focusing on Pakistan as a case study, this work offers practical insights and strategies for deploying AI-driven technologies sustainably in
similar resource-constrained environments and contributes to the broader discourse on the role of AI in global health improvement.

CCS Concepts: • Information systems → Expert systems; • Social and professional topics → Medical technologies; • Human-
centered computing → Accessibility theory, concepts and paradigms;

Keywords: healthcare process optimization, resource-constrained environments, artificial intelligence, sustainable AI integration, pilot
deployment

ACM Reference Format:

Imama Zahoor, Shiza Ihtsham, Muhammad Umar Ramzan, Agha Ali Raza, and Basmaa Ali. 2024. AI-Driven Healthcare Delivery in Pakistan:
A Framework for Systemic Improvement. In ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS '24), July
08--11, 2024, New Delhi, India. ACM, New York, NY, USA 8 Pages. https://doi.org/10.1145/3674829.3675058
(https://doi.org/10.1145/3674829.3675058)

1 INTRODUCTION
Approximately 8 million individuals die in Low- and Middle-Income Countries (LMICs) every year due to
inefficient healthcare practices, with 60% of these deaths stemming from poor quality care [29]. In 2015 alone,
such fatalities resulted in economic losses exceeding US$6 trillion [22]. In Pakistan, there is a higher burden of
disease, ineffective health information exchange (HIE) resulting in resource wastage due to poor allocation and
duplication of efforts, premature and avoidable deaths, and a cycle of poverty stemming from health expenditures
[20]. Additionally, the doctor-to-patient ratio in Pakistan stands at 1:1300, notably below the WHO's suggested
ratio of 1:1000 [19]. This undermines the quality of healthcare by increasing the workload for existing
practitioners and resulting in overburdened and burnt-out healthcare professionals. A major aspect of this
challenge lies in documentation practices in Electronic Health Records (EHRs) and the physician role has become
two-fold: 1) attending to the patient's needs and 2) entering diagnoses, orders, visit notes, and additional
administrative data into the EHR [12].

The proliferation of Artificial Intelligence (AI) in healthcare has emerged as a tool for augmenting physician
expertise and patient care. Recent research has explored its application in LMICs in the following major areas:
clinical decision support systems, treatment planning, screening, triage assistants, and health chatbots [1, 2, 8].
However, most of these applications are separate and standalone systems researched and developed in high-
income countries. This presents an opportunity for a solution that is effective, scalable, seamless to integrate into
the current system, and well-aligned with local norms and health patterns. In this paper, we:

• Identify the foundational gaps in healthcare facilities in Pakistan.

• Introduce Darcheeni, an AI-driven framework designed to fill these gaps and develop an approach to
healthcare delivery in resource-constrained environments.

• Discuss the challanges from the pilot deployment of Darcheeni.

It is important to note that while Darcheeni is formulated as a comprehensive framework based on foundational
research, only a portion of it has been deployed to measure real-world implications and challenges. Therefore, we
refer to Darcheeni as a framework rather than a full working system. Through this short paper, we hope to
contribute to the discourse on AI in healthcare and specifically, provide practical insights and strategies for the
sustainable deployment of such technologies in resource-constrained environments.

2 BACKGROUND
AI is a rapidly evolving field which promises to construct machines that can perform tasks that typically require
human intelligence. It comprises a range of techniques such as Machine Learning (ML), Deep Learning (DL), and
Natural Language Processing (NLP). In particular, AI is poised to transform healthcare by improving the quality,
accessibility, and affordability of care.

2.1 AI in Healthcare

AI can be utilized in several ways to improve healthcare. Traditional Machine Learning techniques provide the
ability to analyze large and complex patient datasets to predict outcomes, optimize medical dosages, and create
personalized treatment plans [16, 32, 36]. They can mitigate the risk of adverse drug effects by predicting them in
advance [26]. Similarly, patients at risk of developing chronic diseases can be identified, leading to faster
interventions and a reduced strain on the healthcare system in the future [30]. Clinicians can be assisted in
making treatment decisions, particularly in predicting therapy responses [16] and in conducting triage for patients
based on urgency, prioritizing high-risk cases, leading to lower waiting times and improved patient flows [11].

Similarly, researchers and practitioners alike have utilized computer vision techniques to analyze medical imagery
and extract relevant diagnostic information. Robust models for disease classification, region segmentation, and
nodule detection are developed for conditions where the relevant data can be collected [38]. Radiology, pathology,
ophthalmology, and dermatology have, in particular, received substantial attention. The use of such methods in
radiology has been so significant that it has quickly grown into its own field of research [7, 27, 33]. These
developments have allowed for a rapid response in times of crisis - for instance in the development and
deployment of detection models for COVID-19 [41].

Previous research has indicated that physicians can dedicate up to 35% of their time towards documentation [17].
Large Language Models (LLMs) can be used to automatically generate such documentation. This can save valuable
time, enabling healthcare professionals to focus on patient care, making the entire process more satisfying for both
physicians and patients. LLMs can also be used to generate a concise summary of a patient's complete medical
history, providing the physician with relevant information and facilitating increased efficiency [23, 35, 39]. They
can also be incorporated into traditional computer-aided diagnosis systems, allowing physicians to ask open-
ended questions regarding specific diagnoses so as to better understand the system.

2.2 AI in LMICs

AI holds the potential for tremendous opportunities for LMICs which are lacking in medical resources, expertise,
and infrastructure [13, 34]. Rural and isolated areas, in particular, may benefit from the introduction of AI-based
health applications. By utilizing software solutions, such applications could dramatically reduce the costs
associated with screening and treatment plan selection for diseases that require specialized expertise or expensive
equipment [5, 9, 18, 31]. Other potential benefits include the reduction of waiting times and provision of a private
channel for those suffering from stigmatizing diseases such as psychiatric pathologies [21, 24]. Maternal and child
health, which are major issues in LMICs, may also find such systems to be particularly useful [14, 37]. Automated
translation services can tailor such solutions for the local language and culture, improving the accessibility and
compliance of such services in areas where cultural beliefs may pose a barrier to healthcare [25].

However, health applications utilizing AI have been developed predominantly within high-income countries and
have only recently begun to be used within LMICs [37]. As such, obtaining an evaluation for these systems within
a local context is a difficult process, often resulting in uninformed decision making within that context [14].
Therefore, it is pertinent to consider the associated risks and challenges that are unique to LMICs. For instance,
the training of AI systems often requires large quantities of high-quality data which is difficult to obtain in LMICs
[34]. Simply relying on data collected from other countries can lead to biased or defective solutions [14].
Applications developed in the context of high-income countries may recommend treatments that are either
inaccessible or prohibitively expensive in low-income countries [37]. There also exists a lack of the expertise and
infrastructure required to create governance models used to guide such technologies, resulting in an adverse effect
on the quality and safety of the system [14]. Such systems may also require a large initial investment to develop,
leading to a lack of deployment in low-income countries. Furthermore, the widespread practice of informal
medicine in certain societies may instead lead to the spread of non-compliant AI applications.

https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 1/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

Research regarding the challenges faced in the practical implementation of healthcare AI systems in LMICs is a
new and burgeoning field. We seek to add to this discourse by providing a comprehensive overview of the
implementation of such a system in Pakistan, where relevant research is particularly scarce. The insights gained
from this implementation may be used to more effectively guide the design and creation of AI-based health
applications in Pakistan and other LMICs.

3 FOUNDATIONAL RESEARCH
3.1 Methodology

The development of Darcheeni involved an investigation of healthcare delivery across a set of 8 outpatient
facilities in Lahore, Pakistan. Our observational study was conducted over a year, from June 2022 to June 2023,
and aimed to identify the gaps in the current healthcare delivery systems that AI could reimagine. The facilities
included:

• Purpose-built Outpatient Departments (OPD) in large tertiary care hospitals,

• Charity-funded clinics,

• Private clinics in upper-middle-class and middle-class, neighborhoods, and

• Purpose-built OPD in large secondary care hospitals.

The observational team consisted of the first, second, and fifth authors. The first author is a computer science
research assistant while the second and fifth authors are medical doctors with experience at the intersection of
technology and healthcare. During the study period, the team accumulated over 200 hours of observation across
the aforementioned facilities. Each observation session lasted anywhere from 2 to 5 hours during clinic hours and
we observed over 100 clinical encounters in total with a diverse demographic of patients in terms of age, gender,
and socio-economic status. Each facility was visited twice and each visit was conducted by at least two researchers
from the team.

Given the sensitivity of the healthcare environment, particularly patient consultations, we did not audio or video
record to respect privacy and confidentiality. Instead, our data collection relied on detailed handwritten field notes
taken by researchers during and immediately after observation sessions. These notes covered aspects such as
patient flow, documentation practices, patient-physician interactions, and the use of medical devices and
technology within the clinical setting. We also conducted informal interviews with healthcare providers, including
physicians, nurses, and administrative staff, alongside patients and their families when possible. These
conversations were instrumental in providing deeper insights into the practical challenges and expectations of the
healthcare delivery ecosystem from those who experience it firsthand. All conversations were paraphrased in our
field notes, with identifiers removed to maintain confidentiality. We then employed axial qualitative coding to
identify broad themes and patterns. By incorporating these personal anecdotes and testimonies into our analysis,
we aim to present a more comprehensive understanding of the healthcare landscape.

We acknowledge some of the limitations of our study. Firstly, it includes a select number of healthcare facilities in
Lahore - a metropolitan city in Pakistan which potentially limits the generalizability of our findings to other, more
rural, regions. Secondly, the reliance on handwritten notes without audiovisual recording may omit non-verbal
cues important for a complete understanding of the healthcare delivery process. Nonetheless, due to the diversity
within the facilities visited, we maintain that our findings present a microcosm of the broader healthcare
challenges faced nationwide and form a foundation for proposing the Darcheeni framework.

3.2 Findings

3.2.1 Patient Identification. Patient identification practices include tracking and managing patient information
across visits, which is critical for ensuring continuity of care and accurate medical history records. At the tertiary-
care hospital, each patient was assigned a new medical record number (MRN) at every visit, lacking retention of
demographic or medical history details. This systemic amnesia led to incomplete patient histories, disjointed care,
and increased morbidity and mortality. A conversation with a family member of a patient revealed that due to the
absence of a consistent medical history, the patient was repeatedly misdiagnosed and treated for the wrong
ailment which led to a deterioration of their actual condition.

While some of the larger institutions have begun integrating the Computerized National Identity Card (CNIC) into
their EHR systems as a means of ensuring consistency in patient identity, smaller healthcare facilities, which
constitute a majority of the accessible healthcare facilities, often have rudimentary practices due to resource
constraints. In these settings, patient identification and record-keeping are frequently managed through paper
records, which are prone to being misplaced or damaged. Moreover, in some instances, patients are only provided
with a slip containing their prescribed medications without any formal documentation of their visit, diagnosis, or
treatment plan. A patient expressed their frustration, stating,

“Every time I visit, it's like starting from zero. Last time, I just got a slip for my meds—no history, no follow-up.”

3.2.2 Overcrowding and Wait Times. Overcrowding was ubiquitous across all sites. The absence of triage and
patient scheduling results in the patients enduring lengthy wait times and little to no distinction between first-
time and follow-up encounters. Patients either self-direct themselves to departments, rely on nurses’ guesses, or
are sent to general practitioners. In one instance, a patient with a broken thigh bone was sent to the emergency
department (ER) after waiting for over an hour in the outpatient department (OPD). Additionally, facilities have
not done a capacity assessment of their infrastructure, personnel, and technical facilities (labs and imaging).
Hence, they cannot plan for timely, humane delivery of care where supply and demand match each other. This
mismanagement not only strains the physical infrastructure but also overburdens the medical and administrative
staff, impacting the overall quality of care delivered.

3.2.3 Substandard Care. Poor delivery of care may be a result:

1. High work pressure on physicians. High work pressure on physicians often leads to rushed consultations, with
doctors seeing 50-150 patients per day and allocating approximately 3 minutes per patient. This prevents
thorough reviews and physical examinations, crucial for accurate diagnosis and treatment planning. For example,
in one of the charity-funded clinics, doctors prioritized speed over depth in patient evaluations, causing them to
overlook critical symptoms or health indicators. Less than half of these oversights were identified after the patient
had left.

2. Lack of accountability. Health outcomes are rarely tracked or directly linked to the interventions suggested by
doctors, creating a gap in quality assurance and improvement mechanisms. There are also little to no mechanisms
in place to hold physicians accountable for their mistakes, leading to carelessness and lapses in diligence on the
part of physicians. For instance, an administrative staff at an OPD revealed that when a patient receives incorrect
medication due to a misdiagnosis, there's often no system in place to address or learn from these mistakes.

3. Poor medical record-keeping. The majority of public setups have no medical records while some use paper
records, resulting in inconsistent and often illegible notes and prescriptions. Most of the current setups delegate
the responsibility of maintaining medical records and histories to the patients. Other setups rely on outdated
electronic systems often containing overlooked or inaccurate information. Furthermore, there is a lack of
communication of care plans to patients and little to no use of exit interviews to explain the treatment or gather
feedback.

3.2.4 Sub-optimal Implementation of Patient Rights. This is categorized by inadequate protection of privacy,
insufficient informed consent processes, and minimal patient engagement in treatment decisions. For instance, in
some clinics, conversations about sensitive health information occur in semi-public areas, compromising patient
privacy. Additionally, patients received treatment without a thorough explanation of the procedures, risks, and
alternatives due to time constraints or staff shortages. For example, most patients in overcrowded facilities were
administered medication without being informed about its side effects or without their explicit consent, as
healthcare providers rushed to see the next in line. This adversely impacts the quality of care, patient trust, and
treatment outcomes. A doctor in one of these settings commented:

“It gets so busy that sometimes we can't provide them [the patients] with all the information. We're doing the best
we can under the circumstances.”

3.2.5 Narrow Use of Technology. The application of technology in healthcare is predominantly geared towards
administrative functions such as billing, inventory management, and scheduling appointments. This focus often
neglects the potential technology holds for enhancing patient care, medical record-keeping, and health outcome
analysis. In recent years, various initiatives to incorporate technology in healthcare practices, such as
implementing EHR systems, have been launched with the intention of streamlining operations and improving
patient care. However, these efforts frequently encounter obstacles due to a lack of infrastructure support,
inadequate training for staff, absence of user research, and resistance to evolve from traditional paper-based
methods. Particularly, adherence to technology-based mechanisms is difficult, with many healthcare providers
reverting to manual processes due to the perceived complexity or unreliability of digital systems. A member of the
IT team in one of the hospitals commented:

“We've tried integrating new software, have so many new ideas, but it's no use if they [the doctors] are not ready to
use it [...] it's more trouble than it's worth.”

There is also a reliance on outdated EHR systems and legacy technologies that are not user-friendly and do not
meet the dynamic needs of modern healthcare delivery. These systems often lack interoperability, making it
difficult to share patient information across different healthcare providers and leading to fragmented care.

4 DESIGN STRATEGIES
Based on the findings from the foundational research, we outlined six goals to address the challenges.

https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 2/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

4.1 Development Goals

The landscape of health systems operating at multiple tiers necessitates system-wide action and foundational
reforms, particularly in Pakistan's context where healthcare challenges are amplified by resource constraints and a
high disease burden. Recognizing that fixes at the micro-level (e.g., healthcare provider or clinic) alone are
unlikely to alter the underlying performance of the entire system. It is also essential to note that AI systems are not
a cure-all for the structural problems faced by healthcare facilities and require a nuanced consolidation of policy
and technology. Keeping this in mind, our proposed framework rests on a multifaceted approach encompassing
the following strategic components:

Figure 1: Development of Local Guidelines in Darcheeni.

4.1.1 Developing Local Standards of Care. Establishing care standards to regional specificities to ensure that
quality care is universally accessible (Figure 1). In regions where healthcare disparities are stark, developing
localized care guidelines can significantly reduce preventable fatalities at a cost sustainable for households and
communities.

4.1.2 Empowering Primary Care. Enhancing the skill set of primary care practitioners in Pakistan not only
optimizes healthcare delivery but also attenuates the long-term costs associated with untreated or poorly managed
conditions.

4.1.3 Expanding Access through Skill Development. Addressing the acute shortage of healthcare professionals in
Pakistan, particularly in rural and under-served areas, necessitates an expansive approach to skill development.
Empowering a broader range of healthcare workers, including nurses, midwives, lady health visitors (LHVs),
pharmacists, and dispensers, to provide care traditionally reserved for general practitioners can significantly
widen the healthcare net, ensuring more individuals receive timely and appropriate care.

4.1.4 Real-Time Health Outcome Tracking. Greater focus on health outcomes by collecting and analyzing real-
time health data. This enables the identification effective interventions and optimization of patient care pathways,
ensuring healthcare delivery is efficient and outcome-focused.

4.1.5 Cultivating Accountability. Implementing real-time data collection to foster a culture of accountability
within Pakistan's healthcare ecosystem not only aids in identifying care delivery lapses and delays but also ensures
healthcare practitioners are cognizant of their roles and responsibilities.

4.1.6 Patient Education for Empowerment. Educating patients about their conditions and treatment options
allows them to actively participate in their healthcare journey, make informed decisions, and advocate for quality
care.

5 ARCHITECTURE OVERVIEW

Figure 2: How Darcheeni works.

To achieve these development goals, we introduce Darcheeni, a framework that leverages AI to streamline
healthcare delivery processes, enhance physician support, and prioritize patient-centered care to address the
challenges of healthcare delivery in Pakistan. The framework consists of five primary modules: holistic data
integration, efficient order processing and care plans, digital profiles, periodic analytics, and streamlined data
transfers.

5.1 Holistic Data Integration

Darcheeni begins by actively listening to doctor-patient interactions, capturing nuances that otherwise might be
overlooked in fast-paced environments. This includes a speech-to-text module that transcribes and translates local
languages to English. Alongside the current interaction, it also aggregates data from diverse sources, including
labs, imaging, and specialist consultations. This ensures the model has a 360-degree view of patient health to offer
holistic and comprehensive advice to physicians by writing a doctor's note and a personalized assessment and plan
for each encounter.

5.2 Efficient Order Processing and Patient-Centric Care Plans

Once the doctor signs off orders, Darcheeni automates the transmission of medication, lab, and imaging orders, as
well as referrals to specialists to streamline the process and lifting responsibilities from the patient. It also sends
personalized care plans to patients’ smartphones to encourage real-time implementation and improve patient
compliance with medications and lifestyle modifications.

5.3 Smart Data Analysis and Digital Profiles

Data from smartphone apps are analyzed by Darcheeni, providing insights for doctors to review during
subsequent appointments. Over time, it constructs digital profiles for each patient, facilitating personalized and
granular care delivery. Similarly, it builds profiles for healthcare providers, including strengths, weaknesses, and
error patterns. These profiles will not only help the model work better alongside the provider by personalizing
each interaction but also help educate and empower both doctors and their patients.

5.4 Periodic Analytics for Continuous Improvement

Darcheeni conducts periodic analyses to create a data-driven, evidence-based healthcare system that is
dynamically responsive to patient needs and healthcare delivery outcomes. It does this by collecting health data
from various sources, including patient interactions, treatment responses, health outcomes, continuous
monitoring, evaluation and adaptive learning. This allows for the identification of trends, patterns, and areas of
improvement in healthcare delivery. By understanding which interventions lead to positive health outcomes and
which areas require enhancement, healthcare providers can implement targeted strategies to improve care quality
and efficiency.

5.5 Efficient Data Transfer

In many LMICs, including Pakistan, the referral process from primary or secondary care to tertiary care often
lacks coordination, efficiency, complete patient information, and accessibility. Through integrated health records
and automated referral pathways, Darcheeni allows for a smoother transition between care levels and anonymized
data transfer between specialists and sub-specialists. It does this by identifying when a patient needs specialist
care and promptly initiating the referral. Additionally, patients’ access to their health records allows them to
follow up on referrals, understand their health conditions better, and make informed choices about their care.

Figure 2 shows how Darcheeni works by completing the loop in the following ways:

1. The doctor and patient conversation is recorded.

2. The AI Model, also titled Darcheeni, pulls patient data from various sources, including lab and imaging reports,
as well as clinical notes from previous encounters from a HIPAA-compliant [4] database.
https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 3/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

3. While the encounter is ongoing, the doctor-facing web application displays real-time processing of the
conversation into discrete problems, pertinent questions to prompt the physician in case of any overlooked
aspects, and the distribution of important sections into bins and packets. The latter is a categorizing methodology,
where bins refer to the general topics of discussion, such as medications, past medical history, and current
problems, while packets refer to each instance of a bin.

4. Once the conversation is over, the doctor stops the recording, and a personalized and editable diagnosis,
assessment, and plan is created. The doctor verifies and signs off on the plan, leading to the creation of three
pathways: billing, storage in the EHR via an API, and transmission to the patient-facing mobile application. The
personalized care plan on the patient-facing application is presented in the form of a smart to-do list, where each
actionable item is sorted by priority.

5. At home, the patient records and updates the patient-facing mobile application. Most phone app entries go to
the medical records and will be reviewed at the next visit. However, some critical entries notify the doctor or clinic
administration for timely intervention.

6 DISCUSSION AND FUTURE WORK

6.1 Pilot Deployment

For the pilot deployment of our system, we collaborated with a tertiary care hospital, established under the
approval of a joint Institutional Review Board (IRB). Within this framework, we constructed a Clinical Innovation
Lab (CIL) in the OPD of the hospital to facilitate a controlled environment for system deployment. Starting in
November 2023, the system's foundational capabilities, note-taking and storing patient records, were
implemented and tested in a real-world clinical setting. This initial phase focused on integrating these core
functionalities into the daily workflows of two doctors operating within the CIL, who collectively attended to an
average of 40 patients per day.

Throughout the deployment, our research team collected over 50 hours of observational data, documenting the
system's performance, user interaction, and overall impact on clinical operations. To ensure ethical compliance
and respect for patient privacy, explicit verbal consent was obtained from all participating patients before
recording their encounters. Any recordings made without consent were immediately discarded, and no identifying
patient information was stored in our records. The setup utilized the existing computers in the OPD and wired
microphones to minimize disruption to the existing clinical infrastructure. This approach not only provided
valuable insights into the system's current capabilities and areas for improvement but also laid the groundwork for
the next phase of deployment. The forthcoming step will involve enhancing the system with advanced
functionalities, including automated diagnosis, assessment, and the creation of personalized care plans.

6.2 Challenges from Pilot Deployment

The initial deployment phase revealed several challenges that we believe to be relevant for any future
implementation of a similar system in an LMIC. These include both technical and usability hurdles that must be
overcome for such a system to be effective:

6.2.1 Transcription Accuracy. Accurate transcription and speaker diarization of the conversation was one such
fundamental hurdle. Any audio transcription system must work well for a combination of local languages and
dialects. In our case, this was a combination of English, Urdu, and Punjabi. Given that Urdu and Punjabi are low-
resource languages, we designed a pipeline to collect and transcribe audio data from multiple clinics to be used to
fine-tune our speech model and improve transcription accuracy. Using such context-relevant training data also
allows the model to become resistant to background noise, which is a prevalent problem in many local clinics.
Transcribing audio files in this way proves to be a slow and expensive process, presenting itself as the main
bottleneck in the fine-tuning pipeline. Languages used within LMICs are often low-resource and so any such
system would require similar fine-tuning before being deployed.

6.2.2 Speed and Scalability. For such a system to be effective, it must function in real time and scale with an
increasing number of users. To achieve the required speed, we utilize a dynamic number of cloud containers that
scale up to match the current number of users. During a conversation, we only transcribe short audio segments,
prioritizing speed. Once a visit is complete, the entire conversation is transcribed once more so as to increase the
final accuracy of the transcription. The full system must ultimately be completely integrated with each hospitals
EHR system. This would, however, be a time-consuming process as each institution has a unique system and
would require a custom solution.

6.2.3 System Usability. The manner in which the system is used also reveals some flaws. As the tool is based
primarily on the audio collected from a conversation, it must rely on the physician to be quite descriptive.
However, we observe that physicians often do not go into the required detail, especially when conducting physical
examinations. Clinics are often overcrowded and noisy, with patients being accompanied by friends or family, who
frequently interject into the conversation. The patients themselves are occasionally too quiet, making their voice
indistinguishable from background noise. To combat this, the team eventually had to implement a "one patient at
a time" policy to ensure high-quality audio recordings. Additionally, early on during the initial deployment, we
realized that physicians must have the ability to alter any part of the output from the system before it is saved. This
would not only increase the accuracy of the output but would also allow us to identify and correct mistakes within
the system in an iterative fashion.

6.3 Potential Impact

Pakistan spends $38 per person per year on healthcare [20]. While increasing this allocation might be
challenging, optimizing its utilization is crucial. In this paper, we outline how Darcheeni has the potential to
deliver quality care, through active listening of doctor-patient interactions, generation of electronic medical
records instantly, and seamless capture of vital information without any additional effort from physicians. This
can enable nationwide accessibility of health records, continuity of care, delivery of quality care, preventive
measures, early intervention, and the effective management of chronic conditions. By doing so, it can significantly
increase the Human Development Index (HDI) and reduce healthcare costs associated with treating advanced and
severe cases. Moreover, automatic record creation for every patient visit establishes a real-time accountability
mechanism for healthcare providers. This holds the potential to reduce healthcare provider absenteeism and
improve patient outcomes.

Similarly, primary healthcare (PHC) facilities in Pakistan are either not accessible or have limited resources [15].
71% of primary care needs are met by general practitioners (GPs), who are not required to complete structured
training and gaps exist in knowledge and skills in clinical practice [10]. Our framework involves the delivery of a
local standard of care so that more patients can get treatment early in the disease, potentially cure or halt its
progression, and consequently take the pressure off tertiary centers. It is also pertinent to note that it takes 17
years on average in the US for new evidence to become medical practice [28]. This measure is infinite in Pakistan
because there is a lack of continuation of medical education for physicians. Upgrading Darcheeni's care guidelines
once every 6 months will ensure that all physicians are treating their patients according to the best guidelines
dictated by international evidence and resource capacity. The availability of reliable and up-to-date health data
will also allow the development of highly targeted and effective health policies.

6.4 Limitations and Future Work

While the Darcheeni framework presents a promising approach to address critical challenges in healthcare
delivery, it is essential to acknowledge several inherent limitations and constraints.

The success of this framework relies heavily on a robust technological infrastructure, requiring electricity, an
internet connection, a web-connected desktop or laptop for the health caregiver, and an internet-connected
smartphone for the patient. While most facilities in the larger cities possess computers and internet connectivity,
the speed of the internet, software versions, and the quality of microphones vary greatly among setups, requiring
significant additional investment at times. As we travel farther away from the metropolitans, the resource
constraints only add up. 62.27% of Pakistanis live in rural regions with limited access to reliable internet
connectivity and advanced healthcare facilities [3]. At the moment, we are being purposeful in seeking
socioeconomic and ethnic diversity as we set up clinical collaborations for the acquisition of our training dataset.

Furthermore, although Pakistan has high smartphone penetration, the gender gap in smartphone (mobile phone
with internet connection) ownership was 43% in 2021, with a gender gap of overall 15% in LMICs [6]. Given
privacy concerns for healthcare data, future research may explore solutions so that women not only have access to
their health data but also have control over access to it. Similarly, since only 22% of the total physicians in
Pakistan serve in the rural areas, where the majority of the population resides, future versions of Darcheeni could
involve allied healthcare workers (AHWs) like nurses, midwives, and compounders who make up the bulk of
primary care providers in the bottom, mostly rural Pakistan and other LMICs [40].

The current deployment of Darcheeni is in its pilot phase, limited to a smaller scale. While this allows for iterative
testing and optimization, extrapolating the findings to a larger, more diverse healthcare ecosystem may pose
challenges such as scalability, affordability, and maintenance costs, which must be addressed to ensure sustainable
integration within the socio-economic landscape of the deployment areas. The research team aims to validate and
refine the model for nationwide deployment by the end of this year. As a future endeavor, scaling the Darcheeni
framework to other LMICs presents an avenue. This would require language adaptation, local health needs
adaptation, affordability, and regulatory compliance. Conducting a longitudinal impact assessment and soliciting
feedback from users, including both healthcare providers and patients, is imperative to evaluate the sustained
effectiveness of Darcheeni.

Currently, the only account of a doctor-patient interaction is recorded by the doctor. Using audio recordings of
encounters shifts the focus from a physician's recall to the patient's narrative. Future work can incorporate the
voice metadata of a patient's initial history to diagnose disease more quickly and help provide better-fitting
therapeutic choices for the patient. It will receive data from patients’ smartphones, analyze it, and share the
insights with the care team. It will also organize the doctor's orders and instructions into a smart list that will
remind the patient via timely notifications to implement the care plan with more compliance. However, we also
https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 4/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

concede that any solution which has access to private data must be carefully evaluated according to local
regulatory laws and strict guardrails must be implemented accordingly.

7 CONCLUSION
In this article, we introduced the architectural design of Darcheeni to the sustainable health research community,
its potential impact on healthcare delivery in LMICs, and the foreseen challenges stemming from resource
constraints. As the field of AI-driven healthcare evolves, we hope to see more technological solutions catering to an
under-represented and diverse user base.

https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 5/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

REFERENCES
[1] [n. d.]. https://www.madiro.org/ (https://www.madiro.org/) Navigate to

[2] [n. d.]. https://ada.com/ (https://ada.com/) Navigate to

[3] [n. d.].

https://www.theglobaleconomy.com/Pakistan/rural_population_percent/#: :text=Rural%2
0population%2C%20percent%20of%20total%20population&text=The%20latest%20value%2
0from%202022%20is%2062.27%20percent
(https://www.theglobaleconomy.com/Pakistan/rural_population_percent/#: :text=Rural%2
0population%2C%20percent%20of%20total%20population&text=The%20latest%20value%2
0from%202022%20is%2062.27%20percent) Navigate to

[4] 1996. Health Insurance Portability and Accountability Act of 1996 (HIPAA).
https://www.hhs.gov/hipaa/index.html (https://www.hhs.gov/hipaa/index.html). Accessed:
2024-03-10. Navigate to

[5] Robert Caprara, Keith L Obstein, Gabriel Scozzarro, Christian Di Natali, Marco Beccani, Douglas R
Morgan, and Pietro Valdastri. 2014. A platform for gastric cancer screening in low-and middle-income
countries. IEEE Transactions on Biomedical Engineering 62, 5 (2014), 1324–1332. Navigate to

[6] Isabelle Carboni, Nadia Jeffrie, Dominica Lindsey, Matthew Shanahan, Claire Sibthorpe, C Butler, et al.
2021. Connected Women-The Mobile Gender Gap Report 2021. GSMA Intelligence: UK 7 (2021).
Navigate to

[7] Garry Choy, Omid Khalilzadeh, Mark Michalski, Synho Do, Anthony E Samir, Oleg S Pianykh, J Raymond
Geis, Pari V Pandharipande, James A Brink, and Keith J Dreyer. 2018. Current applications and future
impact of machine learning in radiology. Radiology 288, 2 (2018), 318–328. Navigate to

[8] Tadeusz Ciecierski-Holmes, Ritvij Singh, Miriam Axt, Stephan Brenner, and Sandra Barteit. 2022.
Artificial intelligence for strengthening healthcare systems in low-and middle-income countries: a
systematic scoping review. npj Digital Medicine 5, 1 (2022), 162. Navigate to

[9] Hugo Jair Escalante, Manuel Montes-y Gómez, Jesús A González, Pilar Gómez-Gil, Leopoldo Altamirano,
Carlos A Reyes, Carolina Reta, and Alejandro Rosales. 2012. Acute leukemia classification by ensemble
particle swarm model selection. Artificial intelligence in medicine 55, 3 (2012), 163–175. Navigate to

[10] Hamida Farazdaq, Jaleed A Gilani, Asra Qureshi, and Unab I Khan. 2022. Needs assessment of general
practitioners in Pakistan: A descriptive cross-sectional survey. Journal of Family Medicine and Primary
Care 11, 12 (2022), 7664–7670. Navigate to

[11] Sabina Ohri Gandhi, Sabina Gandhi, and L Sabik. 2014. Emergency department visit classification using
the NYU algorithm. (2014). Navigate to

[12] Rebekah L Gardner, Emily Cooper, Jacqueline Haskell, Daniel A Harris, Sara Poplau, Philip J Kroth, and
Mark Linzer. 2019. Physician stress and burnout: the impact of health information technology. Journal of
the American Medical Informatics Association 26, 2 (2019), 106–114. Navigate to

[13] Jonathan Guo and Bin Li. 2018. The application of medical artificial intelligence technology in rural areas
of developing countries. Health equity 2, 1 (2018), 174–181. Navigate to

[14] Ahmed Hosny and Hugo JWL Aerts. 2019. Artificial intelligence for global health. Science 366, 6468
(2019), 955–956. Navigate to

[15] Hina Jawaid and Abdul Jalil Khan. [n. d.]. Making Pakistani Primary Care More Resilient. ([n. d.]).
Navigate to

[16] Kevin B Johnson, Wei-Qi Wei, Dilhan Weeraratne, Mark E Frisse, Karl Misulis, Kyu Rhee, Juan Zhao, and
Jane L Snowdon. 2021. Precision medicine, AI, and the future of personalized health care. Clinical and
translational science 14, 1 (2021), 86–93. Navigate to

[17] Erik Joukes, Ameen Abu-Hanna, Ronald Cornet, and Nicolette F de Keizer. 2018. Time spent on dedicated
patient care and documentation tasks before and after the introduction of a structured and standardized
electronic health record. Applied clinical informatics 9, 01 (2018), 046–053. Navigate to

[18] Shivaram Kalyanakrishnan, Rahul Alex Panicker, Sarayu Natarajan, and Shreya Rao. 2018. Opportunities
and challenges for artificial intelligence in India. In Proceedings of the 2018 AAAI/ACM conference on AI,
Ethics, and Society. 164–170. Navigate to

[19] SA Khan. 2019. Situation analysis of health care system of Pakistan: post 18 amendments. Health Care
Current Reviews 7, 3 (2019), 244. Navigate to

[20] Salman J Khan, Muhammad Asif, Sadia Aslam, Wahab J Khan, and Syed A Hamza. 2023. Pakistan's
healthcare system: A review of major challenges and the first comprehensive universal health coverage
initiative. Cureus 15, 9 (2023). Navigate to

[21] Ilona Kickbusch. 2019. Health promotion 4.0., 179–181 pages. Navigate to

[22] Margaret E Kruk, Anna D Gage, Naima T Joseph, Goodarz Danaei, Sebastián García-Saisó, and Joshua A
Salomon. 2018. Mortality due to low-quality health systems in the universal health coverage era: a
systematic analysis of amenable deaths in 137 countries. The Lancet 392, 10160 (2018), 2203–2212.
Navigate to

[23] Chengtai Li, Yiming Zhang, Ying Weng, Boding Wang, and Zhenzhu Li. 2023. Natural language
processing applications for computer-aided diagnosis in oncology. Diagnostics 13, 2 (2023), 286.
Navigate to

[24] David D Luxton. 2014. Artificial intelligence in psychological practice: Current and future applications
and implications.Professional Psychology: Research and Practice 45, 5 (2014), 332. Navigate to

[25] David D Luxton. 2014. Recommendations for the ethical use and design of artificial intelligent care
providers. Artificial intelligence in medicine 62, 1 (2014), 1–10. Navigate to

[26] Guillaume L Martin, Julien Jouganous, Romain Savidan, Axel Bellec, Clément Goehrs, Mehdi Benkebil,
Ghada Miremont, Joëlle Micallef, Francesco Salvo, Antoine Pariente, et al. 2022. Validation of artificial
intelligence to support the automatic coding of patient adverse drug reaction reports, using nationwide
pharmacovigilance data. Drug Safety 45, 5 (2022), 535–548. Navigate to

[27] Maciej A Mazurowski, Mateusz Buda, Ashirbani Saha, and Mustafa R Bashir. 2019. Deep learning in
radiology: An overview of the concepts and a survey of the state of the art with focus on MRI. Journal of
magnetic resonance imaging 49, 4 (2019), 939–954. Navigate to

[28] Zoë Slote Morris, Steven Wooding, and Jonathan Grant. 2011. The answer is 17 years, what is the
question: understanding time lags in translational research. Journal of the royal society of medicine 104, 12
(2011), 510–520. Navigate to

[29] Engineering National Academies of Sciences and Medicine. 2018. Crossing the Global Quality Chasm:
Improving Health Care Worldwide. The National Academies Press, Washington, DC.
https://doi.org/10.17226/25152 (https://doi.org/10.17226/25152) Navigate to

[30] Karin M Nelson, Evelyn T Chang, Donna M Zulman, Lisa V Rubenstein, Freddy D Kirkland, and
Stephan D Fihn. 2019. Using predictive analytics to guide patient care and research in a national health
system. Journal of general internal medicine 34 (2019), 1379–1380. Navigate to

[31] Allisson Dantas Oliveira, Clara Prats, Mateu Espasa, Francesc Zarzuela Serrat, Cristina Montañola Sales,
Aroa Silgado, Daniel Lopez Codina, Mercia Eliane Arruda, Jordi Gomez i Prat, and Jones Albuquerque.
2017. The malaria system microapp: a new, mobile device-based tool for malaria diagnosis. JMIR research
protocols 6, 4 (2017), e6758. Navigate to

[32] Sameer Quazi. 2022. Artificial intelligence and machine learning in precision and genomic medicine.
Medical Oncology 39, 8 (2022), 120. Navigate to

[33] Luca Saba, Mainak Biswas, Venkatanareshbabu Kuppili, Elisa Cuadrado Godia, Harman S Suri,
Damodar Reddy Edla, Tomaž Omerzu, John R Laird, Narendra N Khanna, Sophie Mavrogeni, et al. 2019.
The present and future of deep learning in radiology. European journal of radiology 114 (2019), 14–24.
Navigate to

[34] Laura Sallstrom, Olive Morris, and Halak Mehta. 2019. Artificial intelligence in Africa's healthcare:
ethical considerations. ORF Issue Brief 312 (2019), 1–11. Navigate to

[35] Yiqiu Shen, Laura Heacock, Jonathan Elias, Keith D Hentel, Beatriu Reig, George Shih, and Linda Moy.
2023. ChatGPT and other large language models are double-edged swords., e230163 pages. Navigate to

[36] Murugan Subramanian, Anne Wojtusciszyn, Lucie Favre, Sabri Boughorbel, Jingxuan Shan, Khaled B
Letaief, Nelly Pitteloud, and Lotfi Chouchane. 2020. Precision medicine in the era of artificial intelligence:
implications in chronic disease management. Journal of translational medicine 18 (2020), 1–12.
Navigate to

[37] Brian Wahl, Aline Cossy-Gantner, Stefan Germann, and Nina R Schwalbe. 2018. Artificial intelligence
(AI) and global health: how can AI contribute to health in resource-poor settings?BMJ global health 3, 4
(2018), e000798. Navigate to

[38] Hongtao Xie, Dongbao Yang, Nannan Sun, Zhineng Chen, and Yongdong Zhang. 2019. Automated
pulmonary nodule detection in CT images using deep convolutional neural networks. Pattern recognition
85 (2019), 109–119. Navigate to

[39] Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin
Compas, Cheryl Martin, Anthony B Costa, Mona G Flores, et al. 2022. A large language model for electronic
health records. NPJ digital medicine 5, 1 (2022), 194. Navigate to

[40] Zeshan Zahid, Suleman Atique, Muhammad Hammad Saghir, Iftikhar Ali, Amna Shahid, and Rehan Ali
Malik. 2017. A commentary on telerehabilitation services in Pakistan: current trends and future
possibilities. International journal of telerehabilitation 9, 1 (2017), 71. Navigate to

[41] Jianpeng Zhang, Yutong Xie, Guansong Pang, Zhibin Liao, Johan Verjans, Wenxing Li, Zongji Sun, Jian
He, Yi Li, Chunhua Shen, et al. 2020. Viral pneumonia screening on chest X-rays using confidence-aware
anomaly detection. IEEE transactions on medical imaging 40, 3 (2020), 879–890. Navigate to

https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 6/7
6/5/25, 1:23 PM AI-Driven Healthcare Delivery in Pakistan: A Framework for Systemic Improvement

This work is licensed under a Creative Commons Attribution International 4.0 License
(https://creativecommons.org/licenses/by/4.0/).

COMPASS '24, July 08–11, 2024, New Delhi, India

© 2024 Copyright held by the owner/author(s).

ACM ISBN 979-8-4007-1048-3/24/07.
DOI: https://doi.org/10.1145/3674829.3675058 (https://doi.org/10.1145/3674829.3675058)

https://dlnext.acm.org/doi/fullHtml/10.1145/3674829.3675058?utm_source=chatgpt.com 7/7