Knowledge Vault 2/65 - ICLR 2014-2023
Mihaela van der Schaar ICLR 2020 - Invited Speaker - Machine Learning: Changing the future of healthcare
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

graph LR classDef medicine fill:#f9d4d4, font-weight:bold, font-size:14px; classDef lab fill:#d4f9d4, font-weight:bold, font-size:14px; classDef methods fill:#d4d4f9, font-weight:bold, font-size:14px; classDef models fill:#f9f9d4, font-weight:bold, font-size:14px; classDef transform fill:#f9d4f9, font-weight:bold, font-size:14px; A[Mihaela van der Schaar
ICLR 2020 ] --> B[Machine learning in medicine:
new methods needed 1] A --> C[Van der Schaar Lab:
transforming healthcare with AI 2] C --> D[Breast cancer system: integrates data,
interpretability, forecasting, monitoring 3] A --> E[Automated ML creates clinical
analytics across diseases 4] A --> F[Interpretability, trustworthiness, explainability
crucial for medical ML 5] F --> G[Symbolic metamodels: black-box to
white-box, tailored explanations 6] G --> H[Metamodels: discover hypotheses,
inform clinicians, policymakers 7] A --> I[Disease progression: model history,
risk, disease interactions 8] I --> J[Attentive state-space models: complex,
individualized disease trajectories 9] I --> K[Deep diffusion processes: personalized
comorbidity networks over time 10] A --> L[Personalized screening, monitoring
based on individual risk 11] A --> M[Individualized treatment effects ITE:
personalized causal inference 12] M --> N[ITE theory: rates depend on
response surface complexity 13] M --> O[ITE methods: flexible ML,
bias handling, shared representations 14] M --> P[Non-stationary GPs, ARD priors
for ITE with uncertainty 15] M --> Q[GANITE, counterfactual RNNs for
ITE with many treatments 16] A --> R[ML insights transform healthcare
practice, research, industry 17] R --> S[ML augments clinicians, researchers
with powerful targeted tools 18] S --> T[Clinicians: ML for actionable,
patient-tailored intelligence 19] S --> U[Researchers: ML generates
data-induced hypotheses 20] S --> V[Pharma: ML identifies
individualized treatments 20] A --> W[Effective learning integrates data
with expertise using ML 21] A --> X[Van der Schaar Lab: global
collaboration, empowering ML tools 22] D --> Y[Autoprognosis outperforms current
cancer survival models 23] G --> Z[Symbolic metamodels: interpretable
risk equations, transparency 24] Z --> AA[Inverse metamodels inform
risk reduction strategies 25] J --> AB[Individualized disease trajectories
vs one-size-fits-all 26] L --> AC[Personalized cancer screening
schedules vs preset times 27] M --> AD[ITE moves beyond average
effects to optimize treatments 28] N --> AE[ITE difficulty depends on
response surface complexity 29] R --> AF[ML will augment clinicians,
scientists to improve healthcare 30] class A,B,F,I,L,M,R,W,X,AF medicine; class C,D,E,Y,AA,AB,AC,AD lab; class G,H,J,K,N,O,P,Q,Z,AE methods; class S,T,U,V models; class A transform;


1.-Machine learning for medicine is challenging due to poorly-defined problems, solutions and verification. New problem formulations and methods are needed.

2.-The van der Schaar Lab aims to transform healthcare with new machine learning and AI tools to address complex problems.

3.-A breast cancer decision support system was developed using autoprognosis (integrates data), InVase (interpretability), attentive state-space models (forecasts), and personalized monitoring.

4.-Automated machine learning methods can craft clinical analytics at scale across diseases. Structure kernel learning conquers the curse of dimensionality.

5.-Interpretability, trustworthiness and explainability are key for actionable machine learning models in medicine. Different users need tailored interpretations.

6.-Symbolic metamodels turn black-box models into interpretable white-box models, enabling forward and backward explanations for various users.

7.-Metamodels allow robust discovery of data-induced hypotheses from models. They help clinicians understand predictions and policymakers design screening programs.

8.-Disease progression should be modeled with history, individualized risk, and interactions between diseases, not just Markov models.

9.-Attentive state-space models combine probabilistic hidden Markov models with RNNs to model complex non-stationary individualized disease trajectories.

10.-Dynamic, personalized comorbidity networks are learned using deep diffusion processes to model disease interactions over time.

11.-Personalized screening and monitoring policies should be used instead of one-size-fits-all, based on individual patient risk.

12.-Estimating individualized treatment effects (ITE) moves from average effects to personalized causal inference. It's a complex problem due to unobserved counterfactuals.

13.-A Bayesian nonparametric theory for ITE estimation finds minimax rates depend on response surface complexity (sparsity, smoothness), not selection bias asymptotically.

14.-For large samples, flexible ML tunes hyperparameters; for small samples, selection bias handling and shared representations are key.

15.-Non-stationary Gaussian processes with ARD priors on response surfaces enable ITE estimation with limited data and uncertainty measures.

16.-GANITE uses GANs for ITE estimation with many treatments. Counterfactual recurrent networks optimize treatment plans over time.

17.-Machine learning provides insights on diseases, treatments, prevention, diagnosis, resource allocation, and care pathways to transform healthcare.

18.-We are at the beginning of using ML to augment clinicians and medical researchers with powerful targeted tools.

19.-Clinicians can use co-designed ML tools for actionable intelligence tailored to each patient to improve practice.

20.-Researchers can use ML to generate data-induced hypotheses. Pharma can identify individualized treatments with ML.

21.-Effective learning at scale integrates time series, omics, imaging data with clinical expertise using powerful ML methods.

22.-The van der Schaar Lab collaborates globally with clinicians to provide empowering ML tools and seeks collaborators and students.

23.-Autoprognosis automated ML significantly outperforms current models like PREDICT for breast cancer survival prediction, especially in young and old.

24.-Symbolic metamodels provide interpretable risk equations rivaling autoprognosis' accuracy that meet clinical requirements for transparency.

25.-Inverse metamodel equations can inform patients how to reduce risk through modifiable variables and lifestyle changes.

26.-One-size-fits-all disease progression ignores important factors like genetics. Individualized trajectories are modeled using attentive state space models.

27.-Sending post-op cancer patients for follow-up at preset times ignores individual risk. Personalized screening schedules are needed.

28.-Average treatment effects ignore individual differences. Estimating ITEs moves toward personalized causal inference to optimize treatments.

29.-ITE estimation theory shows problem difficulty depends on treatment response surface complexity. Shared data representations help with limited data.

30.-Machine learning will augment clinicians and scientists to improve healthcare at the individual patient level through collaboration and new tools.

Knowledge Vault built byDavid Vivancos 2024