The End Of Knowledge - Vault 2 - ICLR (2014-2023) - Aisha Walcott-Bryant ICLR 2020

graph LR classDef africa fill:#f9d4d4, font-weight:bold, font-size:14px; classDef ai fill:#d4f9d4, font-weight:bold, font-size:14px; classDef subpopulations fill:#f9f9d4, font-weight:bold, font-size:14px; classDef future fill:#f9d4f9, font-weight:bold, font-size:14px; Main[Aisha Walcott-Bryant
ICLR 2020] --> A[Walcott-Bride: IBM Research Africa
scientist, global impact. 1] A --> B[Africa's AI innovations
for global impact. 30] Main --> C[IBM Research Africa:
partnerships, AI for challenges. 3] C --> D[Africa's diversity enables
complex problem-solving. 4] Main --> E[AI for global health:
disease planning, subpopulations. 5] E --> F[Malaria: deadly in
Africa, but preventable. 6] F --> G[Complex factors in
malaria decision-making. 7] F --> H[AI finds cost-effective
malaria interventions. 8] H --> I[Multi-step RL for
optimal malaria plans. 9] H --> J[Multiple models better
predict malaria futures. 10] F --> K[Malaria AI extended:
RL, optimization, explainability. 11] K --> L[Decision platform for
malaria, scalable to diseases. 12] F --> M[COVID-19 tasks like
malaria intervention planning. 13] M --> N[Malaria AI applies
to COVID-19 pandemic. 14] E --> O[Subpopulations for targeted
health interventions. 15] O --> P[Family planning: unmet
need, discontinuation issues. 16] P --> Q[Mining finds discontinuation
patterns across countries. 17] Q --> R[Results vary: health
concerns, pregnancy, issues. 18] Q --> S[Mining informs analyses,
predictive discontinuation models. 19] Q --> T[Interactive dashboard explores
family planning data. 20] O --> U[Maternal/child health: preventable
deaths in developing world. 21] U --> V[Models find maternal/child
mortality risk factors. 22] V --> W[Nigeria: mortality risks
vary by subpopulation. 23] V --> X[Stratification finds extreme
vulnerability and protection. 24] U --> Y[Next steps: penalties,
interventions, applications. 25] Main --> Z[Dashboards guide targeted
interventions, enable collaboration. 26] Main --> AA[IBM Africa AI:
climate, food, finance, water. 27] Main --> AB[Cross-domain AI insights
from interconnected problems. 28] Main --> AC[IBM AI activities,
demos at ICLR 2020. 29] class A,B africa; class C,D,E,F,G,H,I,J,K,L,M,N ai; class O,P,Q,R,S,T,U,V,W,X,Y subpopulations; class AA,AB,AC future;

Resume:

1.-Ayesha Walcott-Bride, a research scientist at IBM Research Africa in Nairobi, Kenya, discusses Africa's AI innovations for global impact.

2.-Her background includes degrees from Clark Atlanta University and MIT, and projects in Africa that influenced her decision to pursue research abroad.

3.-IBM Research Africa focuses on addressing Africa's grand challenges, transforming society, and having global impact through partnerships and AI/tech.

4.-Africa's rich diversity offers opportunities to address complex problems like optimal energy mix, crisis recovery, climate change impacts on disease spread.

5.-The talk focuses on IBM's AI research for global health, specifically disease intervention planning and characterizing subpopulations to understand health outcomes.

6.-Despite progress, malaria still kills a child every 30 seconds, with 90% of cases in sub-Saharan Africa. Treatment and prevention are possible.

7.-Malaria decision-makers must consider many complex factors and intervention models to reach targets. The intervention space is vast.

8.-IBM uses AI methods well-suited for the complex malaria intervention models to find more cost-effective intervention plans compared to current/expert plans.

9.-Multi-step intervention planning using reinforcement learning finds intervention paths over time that are more cost-effective for reaching malaria prevalence targets.

10.-Combining multiple malaria models provides a better uncertain future description. A case study applies this in Uganda to minimize deaths.

11.-The malaria research has been extended with deep RL, constraint optimization, explainability, and a trusted decision platform for optimal context-relevant intervention plans.

12.-The platform allows collaboration between decision-makers, AI developers and domain experts. It's replicable and scalable to other diseases like COVID-19, HIV/AIDS.

13.-For COVID-19, task forces are using common interventions to contain disease, limit disruption, reduce mortality, constrained by budgets, supplies, health workers.

14.-The malaria control work leveraging complex domain models and AI to explore vast intervention spaces directly applies to the COVID-19 pandemic.

15.-Characterizing subpopulations can enable targeted health interventions for bigger impact. Two examples are given in family planning and maternal/child health.

16.-In family planning, 214+ million women who want to avoid pregnancy are not using modern contraceptives. Discontinuation signals dissatisfaction.

17.-Discriminatory sub-sequence mining extends prefix span to find contraceptive use patterns unique to cohorts who discontinued for different reasons across countries.

18.-Results show common discriminatory sub-sequences for women who discontinued due to health concerns vs other reasons in Kenya, Nigeria, Ghana, Burkina Faso.

19.-Ethiopia lacked discriminatory patterns. Nigeria had data issues. Further analyses reveal patterns for discontinuation due to pregnancy while using contraceptives.

20.-The sub-sequence mining outputs inform causal analyses and predictive discontinuation models. An interactive dashboard explores family planning across countries.

21.-In maternal/child health, 99% of maternal deaths are in developing countries, child mortality is 15x higher in sub-Saharan Africa. Most are preventable.

22.-Predictive models found risk factors for maternal and child mortality as markers for vulnerability. Stratification identifies disproportionately susceptible subpopulations.

23.-In Nigeria, mothers with 3+ births in 5 years outside south-south had 6x increased odds of under-5 mortality (48%) vs 13% average.

24.-Protected populations had 1 birth in 5 years, age under 40, household size 3+, with 1/3 decreased odds of under-5 mortality.

25.-Automatic stratification efficiently explores feature combinations to find extreme vulnerability and protection cases, exploiting optimization function properties for polynomial time.

26.-Next steps: subset complexity penalties, explainable targeted interventions on modifiable factors, clinical applications in MNCH, HIV, COVID-19 susceptible populations.

27.-Interactive dashboards help understand subpopulations and inform targeted interventions. The work requires close collaborations with organizations and academia in Africa.

28.-IBM Research Africa also explores AI for climate, food security, financial inclusion, water access, and novel core AI algorithm development.

29.-These interconnected problems motivate cross-domain AI insights. IBM has several AI activities and demos at ICLR 2020.

30.-Africa's diversity makes it a parallel for AI innovations that can have global impact in understanding subpopulations and disease intervention planning.

Knowledge Vault built byDavid Vivancos 2024