Project Meeting -- 07/01/2019 @ 15:00

[callahantiff]

Meeting Date: 07/01/2019
Topic: Bi-Weekly Meeting
Attendees: @mgkahn

Proposed Agenda:

• Medication verification for Fellow 1 is complete and the second fellow is no longer available. Discuss whether another individual is needed, or if 1 is sufficient.
• Briefly discuss findings I presented at NLM Training Grant Conference
  • Confirm code sets
    • Specifically, the sets that we are investigating in Table 2.
  • Follow-up Meeting with Titus Schleyer to discuss applications of this work in his LHS
• Proposed Outline for Thesis Update Talk (scheduled 07/08/19)
  • Title: Integrating Inferred Molecular Features into a Phenotype Vector Space, with Applications
    • Outline:
      • Computational phenotyping is much more than just genotype-phenotype correlations or finding disease-gene relationships using GWAS. The future of computational phenotyping will have diverse applications including everything from next-generation clinical trial recruitment and precision subphenotyping for improved clinical care to EHR-driven phenotyping in order to support the development of learning health systems.
      • Unfortunately, there are many problems with traditional approaches to computational phenotyping that we would like to overcome, which include:
        1. Generalizability, Translatability, and Scalability (PheKnowVec work)
           • Generalizability - will have some preliminary results to review (similar to those in the NLM slides, but more than ADHD -- in progress)
           • Translatability -- I'd like to talk about the mapping work and the verification (including the survey) that has been done there. I will include the LOINC2HPO paper and
           • Scalability -- will just mention this and with that include how it will be applied to the pediatric asthma mechanisms that Stan has been developing.
        2. Assuming we could address the first problem, the majority of electronic phenotypes, like those developed by the eMERGE consortium, lack inclusion of important biological mechanisms, which is likely a result of the fact that most hospital patients are not currently genetic tested. This is where Med2Mech and PheKnowLator come in. The combined work of these projects aim to produce "right" level for better phenotype identification.
           • PheKnowLator - This is the knowledge graph path embedding work, which aims to imitate molecular mechanisms of human disease. I would like to present the knowledge representation we have been developing. I will mention preliminary results of work I am doing with Ignacio to validate the KG I created (paper being submitted by end of month).
           • Med2Mech - This work while similar to the work we did with HPO using lab tests, but we take it 1 step further going beyond manifestation in an effort to get closer to modeling mechanistic causation. I will present the preliminary results, which were presented at AMIA JS.
• If time, discuss some initial issues with identifying control groups and how queries will need to be in a loop for each code set type.

[mgkahn]

We will meet remotely. I am still in Southern California. M From: "Tiffany J. Callahan" <[email protected]> Reply-To: callahantiff/PheKnowVec <[email protected]> Date: Sunday, June 30, 2019 at 8:24 PM To: callahantiff/PheKnowVec <[email protected]> Cc: "Kahn, Michael" <[email protected]>, Mention <[email protected]> Subject: [callahantiff/PheKnowVec] Project Meeting -- 07/01/2019 @ 15:00 (#98) Meeting Date: 07/01/2019 Topic: Bi-Weekly Meeting Attendees: @mgkahn<https://github.com/mgkahn> Proposed Agenda: * Medication verification for Fellow 1 is complete and the second fellow is no longer available. Discuss whether another individual is needed, or if 1 is sufficient. * Briefly discuss findings I presented at NLM Training Grant Conference<https://docs.google.com/presentation/d/1X-eQegrwdyAy2QP0NCp5Clh47-H0XuzHud2nsSUbXEU/edit#slide=id.g446b88928d_0_0> * Confirm code sets<https://github.com/callahantiff/PheKnowVec/wiki/Initial-Manuscript%3A-Analysis-and-Evaluation-Plan#code-sets> * Specifically, the sets that we are investigating in Table 2. * Follow-up Meeting with Titus Schleyer to discuss applications of this work in his LHS * Proposed Outline for Thesis Update Talk (scheduled 07/08/19) * Title: Integrating Inferred Molecular Features into a Phenotype Vector Space, with Applications * Outline: * Computational phenotyping is much more than just genotype-phenotype correlations or finding disease-gene relationships using GWAS. The future of computational phenotyping will have diverse applications including everything from next-generation clinical trial recruitment and precision subphenotyping for improved clinical care to EHR-driven phenotyping in order to support the development of learning health systems. * Unfortunately, there are many problems with traditional approaches to computational phenotyping that we would like to overcome, which include: * Generalizability, Translatability, and Scalability (PheKnowVec work) * Generalizability - will have some preliminary results to review (similar to those in the NLM slides, but more than ADHD -- in progress) * Translatability -- I'd like to talk about the mapping work and the verification (including the survey) that has been done there. I will include the LOINC2HPO paper and * Scalability -- will just mention this and with that include how it will be applied to the pediatric asthma mechanisms that Stan has been developing. * Assuming we could address the first problem, the majority of electronic phenotypes, like those developed by the eMERGE consortium, lack inclusion of important biological mechanisms, which is likely a result of the fact that most hospital patients are not currently genetic tested. This is where Med2Mech and PheKnowLator come in. The combined work of these projects aim to produce "right" level for better phenotype identification. * PheKnowLator - This is the knowledge graph path embedding work, which aims to imitate molecular mechanisms of human disease. I would like to present the knowledge representation we have been developing. I will mention preliminary results of work I am doing with Ignacio to validate the KG I created (paper being submitted by end of month). * Med2Mech - This work while similar to the work we did with HPO using lab tests, but we take it 1 step further going beyond manifestation in an effort to get closer to modeling mechanistic causation. I will present the preliminary results, which were presented at AMIA JS. * If time, discuss some initial issues with identifying control groups and how queries will need to be in a loop for each code set type. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#98?email_source=notifications&email_token=AA557TSOFJWGWE5DT76GIO3P5FTEHA5CNFSM4H4ODVUKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4G4Q47LQ>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AA557TW5PJNXQ7PMOEPHV23P5FTEHANCNFSM4H4ODVUA>.

[callahantiff]

Meeting Date: 07/01/2019
Topic: Bi-Weekly Meeting
Attendees: @mgkahn
Proposed Agenda:

• Medication verification for Fellow 1 is complete and the second fellow is no longer available. Discuss whether another individual is needed, or if 1 is sufficient.

POST-MEETING UPDATE: Verifying with @trinklek if we need a second reviewer in issue #42.

• Briefly discuss findings I presented at NLM Training Grant Conference
  • Follow-up Meeting with Titus Schleyer to discuss applications of this work in his LHS

POST-MEETING UPDATE: Will make a note to follow-up with Dr. Schleyer after thesis update talk. Would be good to schedule a meeting to follow-up after NLM.

• Proposed Outline for Thesis Update Talk (scheduled 07/08/19)
  • Title: Integrating Inferred Molecular Features into a Phenotype Vector Space, with Applications
    • Outline:
      • Computational phenotyping is much more than just genotype-phenotype correlations or finding disease-gene relationships using GWAS. The future of computational phenotyping will have diverse applications including everything from next-generation clinical trial recruitment and precision subphenotyping for improved clinical care to EHR-driven phenotyping in order to support the development of learning health systems.
      • Unfortunately, there are many problems with traditional approaches to computational phenotyping that we would like to overcome, which include:
        1. Generalizability, Translatability, and Scalability (PheKnowVec work)
           • Generalizability - will have some preliminary results to review (similar to those in the NLM slides, but more than ADHD -- in progress)
           • Translatability -- I'd like to talk about the mapping work and the verification (including the survey) that has been done there. I will include the LOINC2HPO paper and
           • Scalability -- will just mention this and with that include how it will be applied to the pediatric asthma mechanisms that Stan has been developing.
        2. Assuming we could address the first problem, the majority of electronic phenotypes, like those developed by the eMERGE consortium, lack inclusion of important biological mechanisms, which is likely a result of the fact that most hospital patients are not currently genetic tested. This is where Med2Mech and PheKnowLator come in. The combined work of these projects aim to produce "right" level for better phenotype identification.
           • PheKnowLator - This is the knowledge graph path embedding work, which aims to imitate molecular mechanisms of human disease. I would like to present the knowledge representation we have been developing. I will mention preliminary results of work I am doing with Ignacio to validate the KG I created (paper being submitted by end of month).
           • Med2Mech - This work while similar to the work we did with HPO using lab tests, but we take it 1 step further going beyond manifestation in an effort to get closer to modeling mechanistic causation. I will present the preliminary results, which were presented at AMIA JS.

POST-MEETING UPDATE: Went through outline thoroughly. Noted the following changes:

• Need to make sure that it is clear in the presentation that the goal of this work, especially that in 1 above, is to get to a places where these processes and applications can be more generalizable, translatable, and scalable. In order to do that, we have had to implement a lot of manual processes and obtain a lot of manual verification in order to build tools, which will hopefully enable us to become more generalizable, translatable, and scalable.
• We need to note that the lack of genetic data, mentioned in 2 above, is likely NOT the reason that electronic phenotyping efforts like PheKB do not include a measure of biology. We have had knowledge, to varying degrees, of the underlying biological mechanisms driving many diseases. Further, even if we had GWAS and other genetic data available for many or even all patients in an EHR, it is still unlikely prior efforts would have been able to connect the clinical data to the biological data. My work brings in the inference bridge via ontological scaffolding that enables this type of connection.

• [-] If time, discuss some initial issues with identifying control groups and how queries will need to be in a loop for each code set type.

---

Closing this issue. Please re-open if anything above is incorrect or needs additional clarification.