In our pursuit of making quality healthcare accessible and affordable, we were faced with a couple of challenges which include heavy workload on healthcare professionals and poor patient satisfaction. These challenges are not unique to us but to many health practices. With the rise of AI in healthcare, we decided to design interventions that would support healthcare professionals decision making and improve patient satisfaction through mDaktari AI.
mDaktari AI is a clinician assistant that fits into clinical workflows while recognising the bedrock of ethical healthcare practices, thus demonstrating our commitment to responsible AI practices.
At mDaktari, we believe responsible AI is the key to ethical, transparent, and unbiased decision-making, making it accountable like a responsible friend. Our framework is anchored in five guiding principles: fairness, privacy and security, transparency and interpretability, reliability and safety and accountability.
Why Responsible AI Matters
Developing responsible AI models like mDaktari revolves around setting clear and well-informed guardrails throughout the model’s development to ensure patient safety and best clinical practices.
Throughout this process, principles of responsible AI are important because they provide a comprehensive approach that ensure ethical and sustainable practices that are adhered to throughout the model's lifecycle. Responsible AI is the North Star that guides our pursuit of these goals.
Initiatives Undertaken by the mDaktari AI Team
a. Crafting the foundation: Choice of Training Data Sources
Access Afya and the SGHI with clinicians at a focus group session
The genesis of creating an AI model, like mDaktari AI, lies in carefully selecting the appropriate data sources. Our approach includes local context-specific data as well as current international data from reputable organizations such as the World Health Organization (WHO). This helps us to create a responsible and relevant knowledge base for mDaktari AI. We also take into consideration the needs of the end-user in terms of their demographic and socioeconomic status to make data source selection and structuring appropriate and effective.
One of the biggest challenges faced at this stage is the lack of domain-specific expertise among team members. Through Access Afya and Savannah Global Health Institute, we set up a product development team with experts from different specialities like data scientists, medical doctors, software engineers, quality assurance experts, and product managers and owners who continuously contributed their domain knowledge and experience to effectively train the AI model on fairness, privacy and security, transparency and interpretability, reliability, safety, and accountability.
b. Artistry in Structure: Crafting Datasets for Responsible AI Models
Structuring datasets for AI models is not just a technical task; it requires expertise in a specific field. Our approach involves carefully analyzing the main themes in each dataset, combining them into a seamless structure, and making sure that this structure can be replicated and used in similar AI models.
This process is designed to achieve high data quality, directly impacting the performance of our AI models. It allows for feature engineering, enabling us to create customized input features tailored to the specific needs of our AI models. Needless to say, this process can be made more efficient by using existing AI tools such as ChatGPT with close expert human oversight throughout this step. Furthermore, the replicability and reproducibility achieved in structuring datasets play a pivotal role in driving the commercialization of our AI models.
AI Writer Robot by Fanatastic Studio
The often-overlooked art of prompt writing plays a significant role in AI model training. Its purpose is to efficiently retrieve information from your custom dataset and leverage embedded AI algorithms. Therefore, prompt writers have to understand the principles of prompt engineering and the established limitations of AI models.
This gap was evident in the initial prompts used during training the model, and consequently, the AI-generated information appeared inconsistent with the training data used. While it is easy to blame the training dataset for this problem, the role of prompts in this process should not be ignored. To mitigate this issue, the mDaktari team conducted personal and group training sessions on effective prompt writing to develop a model that is ethical, transparent, and unbiased.
The key lesson learned at this stage is that clear, specific, and proofread prompts written in an appropriate format that is understandable to the model and specifies the type of desired response are crucial for generating accurate AI data.
d. Building Trust and Equity: Transparent Telemedicine Integration
As mDaktari AI integrates into telemedicine, transparency emerges as a cornerstone. mDaktari AI takes great care to ensure that healthcare providers and patients have a comprehensive understanding of how AI interventions function within the telemedicine platform. This commitment to transparency fosters trust and comprehension, positioning AI as a valuable aid rather than an enigmatic decision-maker. Additionally, fairness is woven into the system, preventing biases and discrimination and guaranteeing that all patients receive impartial recommendations, regardless of their background.
e. Respect for Individual Rights: Patient Data Privacy
Privacy is another key aspect of the responsible AI approach to establishing mDaktari AI as a responsible steward of healthcare data. We treat patient data with the utmost confidentiality, adhering to privacy regulations, and data access is strictly limited to essential clinical decision-making.
We prioritize patient consent and have established mechanisms that empower healthcare providers to review and override AI recommendations. Our patient-centric approach ensures that human expertise remains central to patient care. We regularly audit our platform to align with responsible AI principles and uphold ethical usage in healthcare.
mDaktari AI is dedicated to enhancing healthcare delivery while unwaveringly upholding the principles of responsible AI, ensuring that patients and providers alike benefit from the integration of AI interventions in a transparent, fair, and privacy-conscious manner.
In conclusion, The mDaktari development team has placed a strong emphasis on the importance of well-informed guardrails throughout the AI model development process. We recommend the formation of cross-disciplinary product development teams, the significance of carefully structured datasets, the need for effective prompts, patient data privacy measures and the importance of continuous monitoring and reinforcement in building responsible AI models.
Access Afya and SGHI team with a patient focus group during a workshop