Leaderboard

Problem Description: In a healthcare claims negotiation & arbitration process, when Out-of-network providers feel that they have been underpaid, they raise negotiation requests. In negotiation stage, claims adjustment team reviews the additional records shared by provider and based on that, they pay extra or stick to the initial payment made. They also explain why the payment amount was decided. If the provider is not happy after that, they raise an arbitration request. The arbitration team selects external arbitrators (from a wide-range of available federal govt-certified arbitrators). The arbitrators are needed to be paid a huge amount (between $500-$1500) by the arbitration team. Arbitrators' decision on the payment amount often goes in favor of the provider. Goal: The goal is to create an AI model, which will be able to suggest a claim payment amount along with an explanation behind selecting the amount in the negotiation stage itself, so that the process can save arbitration fee costs. Historical data from claims adjustment (non-negotiated claims, negotiated claims and arbitrated claims) will be used to build the model. Recommended AI Solution Building Approach: Fine-tuning an existing LLM will be the best fit. Why Fine-Tuning over other available options? Because, this can help us leverage the advanced features of MED-BERT, which is a contextualized embedding model designed for healthcare applications and can be fine-tuned to meet the requirement of the process Below are few key advantages of Fine-Tuning: Efficiency: Less data and computational resources requirement, given the complexity and volume of historical claims data. Performance: When fine-tuned for specific tasks, LLMs can achieve high accuracy and adaptability. Domain Adaptability: The model can be allowed to specialize in the domain, leveraging pre-existing knowledge, while adapting to specific negotiation and arbitration contexts. Speed: Faster to deploy than training a new model from the scratch, even with iterative feedback-based improvements. Limitations of other models: Conventional AI-models: May struggle with the complexity and variability of healthcare claims data. Often lack flexibility and adaptability needed for nuanced decision-making. Lower accuracy and capability of generating detailed explanations. Training a New AI model: Requires extensive computational resources, large datasets. Requires significantly high time and cost investment. High risk of overfitting and poor performance. Flow and Prompt-based Design: May not achieve the required level of accuracy and depth in understanding complex claims data.

When thinking about ways that technology could improve mental healthcare in South Africa, one possible scenario is the use of therapy robots. Which will improve mental health. The Problem: The availability and affordability of mental health services are severely limited for many South Africans, especially those living in rural areas. Due to which it impact severely. Some of the problems with traditional therapy include stigma, long wait times, and an inadequate number of trained therapists. Understanding different cultural contexts and individual needs is crucial for chatbots to be effective in providing initial support, emotional assessments, and useful tools. My recommendation for fixing this issue is to use a broad language model approach that has been fine-tuned. Because of this: By refining a language model, we can better account for South Africa's rich linguistic and cultural diversity. By training on local information, the chatbot can give answers that make sense and are relevant to the situation at hand. Emotional Understanding: Through fine-tuning, the model gains a greater understanding of the nuances of emotional language. This makes it easy to identify indicators of pain and respond appropriately. Rather than building a model from scratch, fine-tuning uses what is already known. This saves time and money. The goal of this method is to save time and resources without lowering the quality. More and more people will be able to get help with their mental health because this approach is easy to adapt to other languages and areas. If we look at alternate Rule-based systems cannot handle the intricacies of human emotions and cultural contexts. Training that starts from beginning requires a large amount of data and computer power, so this is not the greatest method to use it. Flow and prompt-based design can be effective for structured conversations, but they may be unable to give the empathy and insight required for mental health care. Improving an LLM can contribute to the development of a mental health support system that is culturally sensitive, effective, and fulfills the needs of all South Africans.

Scenario: Predicting a Bioequivalence Study for Generic Drug Products In formulation development, a successful bioequivalence (BE) study is critical for a generic drug product. A successful BE study means that the rate and extent of absorption (pharmacokinetics, PK) are statistically equivalent to the reference (brand/innovator) product. Running a human bioequivalence (BE) study can be quite expensive and time-consuming, but there's also a significant risk of failure. This can lead to project delays and affect the overall business case. Key Challenges: 1. Pharmacokinetic variability: The success of a BE study depends on various factors, including the design of the formulation, dissolution data, the drug's in vivo effects, and patient-related factors such as pre-existing conditions and geographical differences. Passing criteria: Regulatory agencies like FDA/Health Canada/EMA require strict statistical passing criteria equivalence (90% CI within 80-125% for AUC & Cmax). An AI that predicts bioequivalence studies could help to customise drug products and reduce the failure probability. A Conventional AI combined with a fine-tuned pharma domain-specific LLM is the best option 1. Conventional AI (First step) Prediction of Bioequivalence parameters requires structured data like formulation properties, dissolution profiles, drug substance data, and/or preclinical PK data. Conventional AI approach includes physiological pharmacokinetic (PBPK) modelling, In Vitro-In Vivo correlation, and statistical equivalence creator. PBPK modelling is the basis for bioequivalence prediction. While conventional AI can deliver precise and consistent results, it also offers regulatory clarity. 2. Fine-Tuned LLM (Second step) A specialized pharma domain LLM can assist in generating reports that meet regulatory standards, extracting relevant study information from existing literature, and outlining a strategy for hypothetical bioequivalence testing. This LLM includes fine-tuning on regulatory guidelines, PK information, and previous pass/fail BE study reports. Why Not Other Approaches? Only Conventional AI: Good for pharmacokinetic data prediction, but lacks explainability and is limited where part of the data is missing. Training from Scratch: Bioequivalence prediction does not need to be developed from scratch, as existing PK knowledge can be fine-tuned. Pure Prompt-Based LLM: Bioequivalence prediction requires precise modelling, so pure prompt-based LLM is not efficient. This approach balances scientific knowledge with AI flexibility (LLM augmentation), making it both innovative, accurate and successful.