WarriorFrontline

Empowering emergency blood availability and equitable Thalassemia education through predictive AI scoring and inclusive multimodal interaction.

Comment

Inspiration

In India, Thalassemia patients often face urgent transfusion needs that demand not just availability of blood, but timely, reliable, and equitable donor response. We observed that while many platforms manage logistics, few solutions predict donor behavior during emergencies — especially using past response data. Additionally, language, disability, and societal discrimination limit access to care for vulnerable Thalassemia patients. I have ideas for the hackathon which align with the highlighted points in the suggested problem statements "Can AI be used for predictive donor availability based on past donation patterns?", "How can we improve patient education on Thalassemia management and care?"

Problem Statement 1: AI-based Prediction of Donor Availability in Emergency Blood Requests

Problem:

In critical care scenarios, particularly for patients with Thalassemia who need recurring blood transfusions, the unpredictability of donor availability remains a major bottleneck. Emergency blood requests often go unanswered or are delayed due to the lack of an intelligent mechanism to prioritize donors most likely to respond. Although donor registries exist (via NGOs like Blood Warriors or government platforms like e-RaktKosh), they are passive repositories without any active behavioral intelligence or real-time ranking mechanisms.

Root Cause:

There is currently no AI-based scoring system that can estimate a registered donor’s likelihood to respond positively to an emergency request. Donor outreach typically happens through broadcast messaging, without consideration of past responsiveness, contextual factors (e.g., time of day), or donor lifestyle variables. This leads to suboptimal matches, wastage of outreach resources, and unnecessary delays in care.

Proposed Approach:

We propose an AI-powered prediction model that computes a donor inclination score (between 0 and 1) for each registered donor, given a new emergency blood request. This score represents how likely the donor is to respond positively, helping NGOs prioritize whom to contact first.

The model is trained using historical records of emergency donation requests and donor responses. Each data instance is represented by contextual and behavioral features engineered at a class-based level (rather than raw values) for generalizability and interpretability. The prediction will happen in near real-time, allowing for tiered messaging, such as automated SMS/email/voice outreach in descending order of likelihood.

Feature Design and Engineering (with Justifications):

To ensure better model generalization and ease of interpretability, we convert continuous or sparse inputs into structured ordinal or nominal classes:

  • Patient Age Group (Ordinal Class): Groups include Infant (0–1), Toddler (1–3), Child (4–12), Teen (13–17), Young Adult (18–25), Adult (26–64), Elderly (65+). Motivation: Donor empathy is influenced by patient vulnerability. Donors are often more responsive to requests for infants or elderly patients.

  • Donor Age Group (Ordinal Class): Includes Young Adult, Adult, Elderly — matching the patient classes starting from 18+. Motivation: Life stage influences motivation, availability, and responsiveness.

  • Distance Between Donor and Emergency Location: Either actual (in km) or classified into Near, Moderate, and Far. Motivation: Geographic proximity directly affects feasibility and urgency of donation.

  • Time of Day (Categorical): Split into behavioral segments: Midnight, Early Morning, Morning, Late Morning, Afternoon, Late Afternoon, Evening, and Night. Motivation: Human alertness and availability vary across the day. A request made at 3 AM may have a different likelihood of response than one made at 6 PM.

  • Donor Professional Class (Nominal): Classes include: Student, Job Seeker, Self-employed, Private Employee, Public Sector Employee, Retired, Homemaker. Motivation: Profession influences routine, mobility, and likelihood to participate. A student may be more flexible than a 9-to-5 employee.

  • Community Affiliation (Binary): Yes/No indicator whether the donor belongs to a social, religious, or activist group. Motivation: Community-driven donors show higher altruistic response rates due to peer influence and group messaging.

  • Required Blood Group (Categorical): One-hot encoded for A⁺, A⁻, B⁺, B⁻, AB⁺, AB⁻, O⁺, O⁻. Motivation: Certain groups (e.g., O⁻, AB⁻) are rare and harder to fulfill, affecting urgency and matching probabilities.

  • Quantity of Blood Required (Categorical): Binned into Low, Medium, High. Motivation: Higher volume requests may discourage casual donors; modeling this helps rank them accordingly.

The AI model (Logistic Regression, LightGBM, or even Transformer-based classifier for scalability) will learn donor behavior patterns based on previous responses and current context. The predicted score will guide priority-based outreach, reducing wait times and improving patient survival chances.

Problem Statement 2: Inclusive Thalassemia Patient Education using Multimodal Foundation Models

Problem:

Thalassemia patients often lack access to comprehensible, inclusive, and accessible educational resources regarding their lifelong treatment, transfusion protocols, dietary restrictions, and psychosocial support. Many patients come from rural areas, diverse linguistic backgrounds, or marginalized social groups (e.g., transgender, neurodivergent), and are either unable to understand mainstream content or hesitant to ask for help due to stigma or lack of safe space.

Root Cause:

Current patient education platforms are monolingual, text-heavy, and non-inclusive, assuming literacy, digital fluency, and a standard socioeconomic background. Moreover, there's no mechanism for anonymous peer-to-peer sharing, no assistive tools for disabled users, and no personalized way to keep patients engaged.

Proposed Approach:

We propose a Multimodal Thalassemia Education and Support Platform powered by foundational models that deliver inclusive, accessible care education to all patients, regardless of location, literacy, disability, or social background.

The System Comprises:

  • Multilingual Support via real-time **Text to Text translation Motivation: Overcomes regional language barriers (e.g., Hindi, Telugu, Marathi, Bengali), ensuring that patients from diverse linguistic backgrounds can understand educational content in their native language.

  • Speech-to-Text (STT) and Text-to-Speech (TTS) modules Motivation: Enables interaction for visually impaired patients or individuals with low literacy by allowing them to listen to or speak queries and content.

  • Screen Reader & Accessibility Optimization (including auditory navigation, keyboard operability, and alt-text tagging) Motivation: Empowers blind and partially-sighted users to navigate and consume platform content through voice-driven or auditory interaction.

  • Anonymized Discussion Forums modeled after pseudonymous platforms, where users from marginalized or discriminated communities (e.g., transgender individuals, those with autism) are assigned pseudo-identities Motivation: Builds a safe, stigma-free, and credibility-preserving environment that encourages honest participation and knowledge-sharing.

  • Multimodal Chatbot powered by a knowledge retrieval system trained on validated medical content and community discussions. Accepts both voice and text inputs and returns responses in the preferred language or modality. Motivation: Offers personalized, context-aware answers and supports both verbal and textual interaction, reducing misinformation and enhancing user trust.

  • Weekly or Fortnightly Digest Generator that compiles audio and regionally translated summary documents or FAQs Motivation: Helps patients stay regularly informed in a light, accessible format without overwhelming them with complex medical jargon or excessive information.

Built With

Share this project:

Updates