Making AI-Enabled Evidence Work for Pakistan
Author: Dr Samreen Sadiq1,2
1. Ghurki Trust Teaching Hospital (GTTH)
2. Ghurki Journal of Medical and Health Sciences
Imagine recovering from a knee replacement in a village two hours from the nearest hospital. There is no follow-up physiotherapist to check your progress and no easy way to know whether your recovery is on track. For many patients across Pakistan, this is the daily reality of post-surgical rehabilitation in a low-resource setting.
Closing that distance, between the operating theatre and any structured follow-up at all, is what our telerehabilitation research has been working towards for the past several years. It is also what has led our team at the Center of Clinical Best Practices (CCBP), Ghurki Trust Teaching Hospital, under the visionary leadership of Prof Dr Amer Aziz, to artificial intelligence (AI), which we now use in building evidence for our country’s health system.
From Telerehabilitation to AI: A Natural Extension
To give some background, our doctoral research centred on telerehabilitation: remotely monitoring patients’ recovery after total knee arthroplasty (TKA), a common surgery for advanced knee arthritis. Through that work, we introduced structured remote monitoring, systematic data collection at fixed intervals and consistent tracking of the same patients over months. Longitudinal data gathered in such ways are what any model needs to be useful, and though rehabilitation services in our setting usually do not have such data at their disposal, when it can be collected, AI offers opportunities to improve patients’ post-surgical rehabilitation. For instance, AI may perform pattern recognition across the dataset, so that a patient whose gains have flattened, or whose pain has risen and adherence fallen, reaches their physiotherapist before the next scheduled contact.
To be clear, AI can’t replace the physiotherapist. In the monitoring period since our telerehabilitation research, a physiotherapist has been reviewing each patient at regular intervals.
AI Has Already Changed How We Work
During the doctoral research, our first full guideline adolopment was slowed by the small core team performing the review, extraction and consensus work. In fact, we leaned on supervised trainees to get through the literature at all. Yet, recently, that bottleneck in our practice has been removed by AI. The CCBP now runs systematic review protocols written specifically for guideline contextualisation, with title-and-abstract screening conducted with AI support, whereby human reviewers verify every decision and resolve all disagreements. Extraction into structured evidence tables then proceeds in the same way, with the evidence checked line by line against source papers. In this approach, the judgement, risk-of-bias assessment, certainty ratings and recommendations themselves remain in the hands of humans, and a multidisciplinary panel still meets in a room to debate the Evidence-to-Decision tables.
What AI has given us is new capacity. Thanks to its support, a centre of our size is now running guideline adolopment in parallel across medicine, surgery, paediatrics, orthopaedics, obstetrics and gynaecology, cardiology, ophthalmology and anaesthesia. If we were still resourced by reviewers alone, we would have been choosing one speciality and asking the rest to wait.
This is why World EBHC Day 2026’s theme, Evidence and AI: People at the Centre, matters to us. AI holds real promise for evidence-based healthcare, but we wish to stress that its benefits will only be realised if this technology is governed carefully and global evidence adapted deliberately.
Why ‘Adopting’ Global Evidence Is Not Enough
A guideline written for a well-resourced country cannot simply be applied in Pakistan and expected to work the same way. Instead, it must be adapted to our infrastructure, patients, resources and patterns of utilising services. GRADE-ADOLOPMENT is a structured method for performing this work: adopting, adapting or developing new recommendations. Its support is crucial in resource-constrained environments likes ours at the CCBP.
The Process of Revising a Guideline
The adolopment performed in our doctoral research led to the development of a clinical practice guideline and referral pathway specific to Pakistan for rehabilitation following TKA. In the process, following systematic search and appraisal, the American Physical Therapy Association (APTA) guideline on physical therapists’ management of TKA was selected as our single source, as it was the most methodologically sounds and used GRADE transparently. Sixteen recommendations were extracted verbatim, then reviewed one at a time against our setting, and sorted into adopt, adapt or exclude. Ten were adopted, five adapted through Evidence-to-Decision tables and one excluded.
APTA recommends supervised physical therapy following TKA to optimise outcomes. We did not question the evidence, but we questioned the assumption behind it, that is, that supervision means a patient being physically present in a clinic. In Pakistan, there is a disparity in access to specialist rehabilitation between urban tertiary hospitals and peripheral facilities, with the patient and/or a companion required to cover travel, and physiotherapist staffing is very variable. If adopted unchanged, the recommendation would have matched practice in Lahore and yet been unachievable almost everywhere else. Using an Evidence-to-Decision table, we rewrote the guideline: structured, supervised physical therapy should be provided following TKA, or where frequent in-person supervision is not feasible, a hybrid supervised and home-based model may be used. The clinical intent was untouched; what changed was the delivery assumption.
Plus, we updated the strength of the guideline for the context of Pakistan, which was set as conditional at moderate certainty. We also removed continuous passive motion altogether, on feasibility grounds, as the equipment and the staffing to run it are not reliably present in the settings this guideline is meant to serve. Leaving it in would have produced a recommendation that documented our aspirations rather than our practice.
Furthermore, we added five recommendations following best-evidence review and expert consensus: safe toilet use, including avoiding squatting and using a raised seat or commode; family caregivers written into the rehabilitation program as participants rather than left outside it; structured home-based rehabilitation with scheduled telephone or telerehabilitation follow-up for patients discharged early or living remotely; dietary counselling; and culturally and linguistically appropriate patient education. None of those could have been discerned by reading the source guideline more carefully. Additionally, they were accompanied by guidance on modified prayer (Namaz) positions that avoid deep knee flexion in early recovery. We could only suggest these positions, not recommend them, because of the low certainty of evidence. This was clearly stated in the guideline. We attribute the lack of evidence to limited interest in researching prayer positions in the United States and Europe, but we felt this guidance we important to include, as how best to pray after a knee replacement is one of the first things our patients ask.
The People Who Get Left Behind
Let’s be clear: the development of this adapted guideline for Pakistan has not solved all problems overnight. In our experience, the key barriers to effective follow-up, which cut across generations, are internet bandwidth, cost, competing work and unfamiliarity with remote assessment, and those persist today. In particular, rural patients cannot access telerehabilitation if they lack stable video connectivity. As a result, data on much of this population never reach the evidence base. That is concerning, because if the evidence ecosystems that inform AI cover only the digitally connected, we will widen the inequities that evidence-based healthcare exists to close.
We should be honest about our own failure in this regard to source evidence from those being left behind. In developing the adapted TKA guideline for Pakistan, patient and caregiver perspectives were captured largely through expert consensus, as an interim surrogate, because direct engagement across a diverse population would have delayed the work. We recorded this as a limitation, and it is the first thing we intend to fix. In the coming years, AI-supported evidence tools need to be designed with underserved populations from the outset, otherwise the ‘people at the centre’ of this year’s theme will become the people left at the margins.
Lessons Learned
Here are the key things we have learned, sometimes the hard way:
- What worked: AI screened and extracted, while GRADE-ADOLOPMENT and a clinical panel decided, which gave clinicians across our departments the confidence to use the results
- What we would do differently: Future developments will involve frontline clinicians, patients and caregivers earlier – before the guideline has been finalised – as contextualisation works best as a starting principle
Putting People Back at the Centre
For AI and evidence to genuinely serve people rather than sideline them, the tools must be locally adapted, the process must be transparent and human clinical judgement must remain firmly in the loop. To return to the patient we began with, two hours from the hospital: what reaches her is the delivery of a guideline that clinicians in Lahore took apart and rebuilt around the patient’s connectivity and her needs for care and prayer. That was achieved by using AI to screen the literature and draft evidence tables, which it can complete so quickly that we are able to do the same across several specialties at once. Nonetheless, throughout the delivery of that adapted guideline in the patient’s follow-up, a physiotherapist is still the one making the calls. From where we sit, that is what it looks like to put people at the centre.
Key Messages
1. Adaptation is not optional: global evidence and guidelines must undergo a structured local adaptation process, such as GRADE-ADOLOPMENT
2. Institutional infrastructure matters: a dedicated unit for evidence adaptation may be introduced with the capacity freed up through the adoption of AI
3. Equity must be designed in from the start: AI-enabled evidence synthesis tools must be fed data that equitably represents digitally excluded and rural populations
References
Arshad, A., Zeeshan, M. H., Mustafa, M. A., Rizvi, N., Rehman, A. A., Pervez, A., & Haider, A. H. (2024). Implementing GRADE-ADOLOPMENT for clinical practice guidelines in resource-limited settings: The AKU experience. Proceedings of the Pakistan Academy of Sciences: B. Life and Environmental Sciences, 61(S), 63–68. https://doi.org/10.53560/PPASB(61-sp1)1009
Jette, D. U., Hunter, S. J., Burkett, L., Langham, B., Logerstedt, D. S., Piuzzi, N. S., Poirier, N. M., Radach, L. J. L., Ritter, J. E., Scalzitti, D. A., Stevens-Lapsley, J. E., Tompkins, J., Zeni Jr, J., & American Physical Therapy Association. (2020). Physical therapist management of total knee arthroplasty. Physical Therapy, 100(9), 1603–1631. https://doi.org/10.1093/ptj/pzaa099
Schünemann, H. J., Wiercioch, W., Brozek, J., Etxeandia-Ikobaltzeta, I., Mustafa, R. A., Manja, V., Brignardello-Petersen, R., Neumann, I., Falavigna, M., Alhazzani, W., Santesso, N., Zhang, Y., Meerpohl, J. J., Morgan, R. L., Rochwerg, B., Darzi, A., Rojas, M. X., Carrasco-Labra, A., … Akl, E. A. (2017). GRADE Evidence to Decision (EtD) frameworks for adoption, adaptation, and de novo development of trustworthy recommendations: GRADE-ADOLOPMENT. Journal of Clinical Epidemiology, 81, 101–110. https://doi.org/10.1016/j.jclinepi.2016.09.009
To link to this article - DOI: https://doi.org/10.70253/OIYO7540
Disclaimer
The views expressed in this World EBHC Day Blog, as well as any errors or omissions, are the sole responsibility of the author and do not represent the views of the World EBHC Day Steering Committee, Official Partners or Sponsors; nor does it imply endorsement by the aforementioned parties.
