Knowledge Vault 5 /62 - CVPR 2021
Artificial Intelligence for Global Health
Pablo Arbeláez
< Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef symphonia fill:#f9d4d4, font-weight:bold, font-size:14px classDef responsibleAI fill:#d4f9d4, font-weight:bold, font-size:14px classDef smartpooling fill:#d4d4f9, font-weight:bold, font-size:14px classDef medicalAI fill:#f9f9d4, font-weight:bold, font-size:14px classDef drugdiscovery fill:#f9d4f9, font-weight:bold, font-size:14px A[Artificial Intelligence for
Global Health] --> B[SYMPHONIA: responsible AI research
center in Colombia. 1] B --> C[Research: vision, sustainability, health. 2] B --> D[Vision for drug discovery,
molecular interactions. 3] D --> E[Adversarial training improves
interpretability, focuses substructures. 4] D --> F[Models identified potential
FPPS antiviral molecule. 5] B --> G[Generic medical image
segmentation: organs, lesions. 6] G --> H[Challenges: annotations, computation,
generalizability. 7] G --> I[Robustness, interpretability vital
for medical AI trust. 8] G --> J[Segmentation model: competitive,
robust to noise. 9] B --> K[CHD risk in infants,
early AI diagnosis. 10] K --> L[TERFOS: CHD screening,
treatment in Colombia. 11] K --> M[UltraGAN: enhances echocardiograms,
preserves anatomy. 12] M --> N[Enhanced images improve
heart segmentation. 13] A --> O[COVID-19: 160M+ cases,
3M+ deaths mid-2021. 14] O --> P[PCR testing critical
but limited. 15] O --> Q[Two-step pooling efficient
at low prevalence. 16] Q --> R[Two-step inefficient as
prevalence rises. 17] O --> S[Smart pooling: ML
predicts status, optimizes. 18] S --> T[App collects data,
informs algorithm. 19] S --> U[Learns from results,
adapts to changes. 20] S --> V[Proof-of-concept: 2K tests,
$100K saved. 21] S --> W[Deployed amid high
ICU occupancy, positivity. 22] S --> X[Historical data trains
robust models. 23] S --> Y[1.5x efficiency over
individual at 35%+. 24] S --> Z[Benefits vs two-step
at high prevalence. 25] S --> AA[Enhances PCR, deployed
successfully. 26] S --> AB[Robust to data,
outperforms pooling. 27] S --> AC[Interpretability, optimizes lab
resources. 28] S --> AD[Generalizes to other
pooled screening diseases. 29] S --> AE[Responsible AI for
social good. 30] class A,B symphonia class C,H,I,J,K,L,M,N medicalAI class D,E,F drugdiscovery class G,O,P,Q,R,S,T,U,V,W,X,Y,Z,AA,AB,AC,AD,AE smartpooling


1.- SYMPHONIA is an AI research center at Universidad de Los Andes in Bogotá, Colombia, focused on responsible research and ethical principles.

2.- Key research areas are computer vision, sustainable development, and global health applications.

3.- They apply computer vision techniques to drug discovery to predict molecular interactions and identify promising pharmaceutical candidates.

4.- Adversarial training improves interpretability of drug discovery models by focusing on relevant molecular substructures.

5.- Their AI models identified a potential antiviral molecule active against the FPPS protein, demonstrating the power of in silico screening.

6.- Generic medical image segmentation aims to identify organs and lesions in 3D diagnostic scans like X-rays, MRIs, CTs.

7.- Challenges include limited annotations, computational demands of 3D volumes, and need for methods that generalize across tasks.

8.- Adversarial robustness and interpretability are important for trust in medical AI systems, which can be vulnerable to attacks.

9.- Their generic medical segmentation model obtains competitive performance across tasks and is more robust to adversarial noise.

10.- Congenital heart diseases (CHDs) put infants at risk when not detected until birth. Early diagnosis with AI can help.

11.- Their TERFOS project seeks to implement a national CHD screening and treatment program in Colombia.

12.- Echocardiogram image quality affects CHD assessment. Their UltraGAN method enhances image quality while preserving anatomical structures.

13.- Enhanced images improve downstream segmentation of heart chambers and structures.

14.- The current COVID-19 pandemic has caused over 160 million cases and 3 million deaths worldwide as of mid-2021.

15.- Molecular PCR testing is critical but costly. Personnel and supply shortages limit massive testing.

16.- Two-step pooled testing increases efficiency at low disease prevalence by testing sample pools, then positive pools individually.

17.- As disease prevalence rises, two-step pooling becomes inefficient since most pools test positive.

18.- Smart pooling uses machine learning to predict patient COVID-19 status and optimize pooling, maintaining efficiency at high prevalence.

19.- A mobile app collects patient data to inform the smart pooling algorithm. Optimized pools are tested at the lab.

20.- Continuous learning from test results allows the model to adapt to changing data distributions over time.

21.- A proof-of-concept saved over 2,000 tests worth $100,000+ in 2 weeks, enabling expanded free testing for at-risk populations.

22.- Large-scale deployment in Bogotá amid 95%+ ICU occupancy and 30%+ positivity rate, in partnership with the city and NGOs.

23.- Historical data from millions of patients used to train more robust smart pooling models.

24.- Smart pooling maintains 1.5x efficiency gains vs individual testing even at 35%+ positivity rate with optimal pool size.

25.- The approach provides significant benefits over two-step pooling at high prevalences seen currently.

26.- Smart pooling enhances rather than replaces PCR testing and has been successfully deployed in the field.

27.- The method is robust to different data types and populations, consistently outperforming conventional pooling.

28.- Smart pooling provides greater interpretability to PCR results and optimizes lab testing resources.

29.- The technique generalizes beyond COVID-19 to other diseases where pooled lab screening is used.

30.- This demonstrates the responsible use of AI for social good. The team invites collaborations to deploy smart pooling more widely.

Knowledge Vault built byDavid Vivancos 2024