Thank you for being here, and participating in this process. I have broad research intersts, which is challneging to talk about in 15 minutes. Hence I will stick to my broad vision, what I set out to achieve, where I am in that path (results), and where I intend to go in future.

Vision

  1. I am interested in getting actionable knowledge from data, particularly in domains where the underlying physics, biology or pathlobiology might not be completely understood yet. Examples might be: predicting case counts in a pandemic, or predicting future mutations and emergence of new pathogens, forecasting when and where for crime in cities, estimating future risk fo complex diseases.

  2. I am interested in getting actionable insights in complex systems, that help clinical decisions, accelerate science and distill new theory. In SOA tools like ML often we need to begin with good informative features, which is not possible in domaind where we lack good understanding, and hence my goal is to push this notion of zero knowledge discovery.

Journey

  1. In the journey so far, I have identified/developed three core ideas or building blocks or algorithmic principles, that then spawn solutions to diverse problems, important and intersting enough to attract significant external funding.

  2. Briefly these three distinct building bloacks define a new approach to EHR analytics, a new approach to analyze complex multivaraite dependencies (the Q-Net framework), and approach that fundementally rethinks stochastic time series modeling.

People

  1. These are all the associates and students I have worked with. Yi and Nick are my postdocs (Yi just left for BNL, both are math PhDs), and teh rest are graduate or UG students in CS and math.
  2. Here are my collaborators across the campus and beyond. And showing the diversity of applications I have engaged, here are all the home departments of my collaborators outside medicine. psychiatry, pediatrics, cardiology, pulomonary care, statistics, PME, social sciences.

EHR

  1. Lets talk about the first category of problems: diagnosis and screening of complex diseases with electronics health records. I designed a new approach to leverage comorbidities effectively to estimate the future risk of a disease, I call these deisease specific estimates xCoR, where x specifies the disease.

  2. Universality. The diseases we have considered so far include asd ad ipf ; These are with the exception of ASDD are fatal diseases with no good screening tool that may be administerd in primary care, leading to missed and late diagnoses. All serious dieases have co-morbidities, they generate fingerprints in our medical histories. Can we use teh vast patianet history data that we have access now, to find and use these patterns?

  3. Conventional ML -- which aims to mimic the physician -- will not do. For these diseases we dont not know enough of the pathobiology to pre-specifiy all risk factors, instead we leverage underutilized diagnostic moadlities -- complex historical patterns -- to complement the physician's experience.

  4. ASD: problem is delayed diagnosis which delays interventions

  5. We know children with autism have higher rates of many diseases. But the hetergenity of autism makes it non-trivial to naively use these patterns.

  6. Our approach cuts down teh false posituves by half.

  7. And here is the ROC, and this (MCHAT) is the state of teh art screening.

  8. And we get to close to 90% AUC at 4 years.

  9. This is the inferred comorbidity spectra: showing how individual diagnoses module autism risk.

  10. How does this work? we infer generative models of stochastic proceses that drive the diagnostic histories, making possible to leverage longitudinal effects beyond just binary presensce/absense of risk factors.
  11. These generative models are known as probabilistic automaata: is an example of PFSA. Theer are subtle differences in teh models inferred in teh control and treatment categories, which are then quantified and used to optimize our predictions.
  12. The algorithm uses 100s of such models pairs to respresent how the diagnosis patterns differ between the control and the treatment sets.
  13. BP, we aim to predict if a patient with depression will experience a manic switch.
  14. IPF. 3-5 years survival post-diagnosis. We get AUCs approaching 88%, and can boost survival from ~100 weeks to >200 weeks on average, due to earlier diagnosis. Disease modifying drugs exist now, so this can be a big deal.
  15. ADRD. 67% to 87%. And we can predict AD diagnosis upto a decade in future.
  16. Teh list if disorders we can target are expanding every week, e.g. cardiac risk in perioperative HK surgery, we estimate the adoption of our tool can save ~10K lives over teh next 20 years.
  17. Prospective trials under way right now
  18. new approach to universal screening across the human disease spectrum

QNET

  1. QNet Architecture: originally developed to predict evolutionary trajectories of emerging viruses. But is actually a general framework to model complex multi-variate systems.
  2. How do we compute distance or similarity between viral strains.. edit distance or the number of mutations they are apart do not reflect biology. The Qnet allows us to infere dependencies between mutations, and induces a biology aware distance metric the qdistance, which adapts to the organism, the evolutionary processes, the selection pressures etc. Using the qdistance we can estimate if two sequences are close: in the sense if one is likely to jump to another in the wild.
  3. example of qnet construction
  4. example that we can indeed predict fututre muations. We outperform WHO in 98% of next-flu-season strain prediction over the past 20 years.
  5. And it works in experiment as well (Aaron's lab). We predicted variants on the SARS-CoV2 spike protein uisng Qnets, and a random sample from those were evaluated in the lab: all of them expressed, were functional ie bound to teh human ACE2 receptor, and soem evaded standard neutralization.

Qnet future

  1. Future: NORAD for Biological/pandemic threats: rand-ordering strains in teh animal reservoirs and targeting them before teh first human infection. ANd if a pandemic still happens then predict timelines, and variants, and help design escape resistant vaccines.
  2. As I said Qnets are applicable to any complex multi-varauate systems, and the second application I am looking into with Erika Claud is gut microbiome modeling for infants.
  3. Model ecosystem evolution, and forecast in individual infants how their microbiome evolves, and matures.
  4. And we are immesely successful, as you can see.
  5. Future, map to disease states, and optimize interventions for microbiome tuning.
  6. Application to belief shifts
  7. examples: opinions or beliefs have dependencies, and we can learn them using qnets from survey data (like the GSS), and this model then allows us to make predictions on how a particualr individual is going to respond to a question, before it is asked.
  8. Social experiment (collab with James Evans) to validate
  9. Also since we can predict responses, this opens up the possiblity of an algoithmic lie detector. We can tell if someone;s responses are consistent with their past responses using the complex inferred models, and is impossible to cheat, and will have immense applications in psychometric testing, and general survey intergrity determination.

Time Series

  1. New approach to modeling stochastic phenomena, that addresses some major gaps in current state of the art. One application is using past flu incidence data to predeict COVID case count.
  2. And in teh CDC backed forecast hub, we beat everyone, including stanford, mit, columbia, LANL .
  3. SImialr underlying principles allow us to model rare and extreme pehnomnea, which is challenging even for state of teh art ML, including extereme weather evenst, earthqaukes, crime.
  4. without going into details, this capability arises from identifying emergent self-simialrities in stoichasti system that produces compact models (order of magnitude smaller that deep learning frameworks, leading to superior predictive performance)
  5. In crime, particualrly, we can also use our generative models to look for enforecemnt bias. So the predictive ability both empowres the state, and opens up new ways to audit effect of state policy decisions.
  6. We would have won in 119 out of 120 categories.
  7. few mathematical insights to diver applications with nontrial impact.

Future

Table of Contents