commit 1cbf160c7e9fd7d39077892a5b547256ad9a26f9
parent bbe04c2d5afdcdd515606605766dfa91a675e5f0
Author: C. Beau Hilton <cbeauhilton@gmail.com>
Date: Thu, 27 Aug 2020 09:02:56 -0500
memex update
Diffstat:
4 files changed, 173 insertions(+), 8 deletions(-)
diff --git a/index.md b/index.md
@@ -71,6 +71,7 @@ Vannevar Bush coined the term "memex" for a system of extending or indexing one'
- [anki](learn/anki.md)
- [memory](learn/memory.md)
+- [want-from-teachers](learn/want-from-teachers.md)
## lectures
### 2019
@@ -79,8 +80,9 @@ Vannevar Bush coined the term "memex" for a system of extending or indexing one'
### 2020
- [2020-07-02-noon-conference](learn/2020-07-02-noon-conference.md)
- [2020-07-06-noon-conference](learn/2020-07-06-noon-conference.md)
+- [2020-08-27-grand-rounds](learn/2020-08-27-grand-rounds.md)
-## blogs
+## blogs (good ones)
- [Olga Botvinnik](learn/olga_botvinnik.md)
diff --git a/learn/2020-08-27-grand-rounds.md b/learn/2020-08-27-grand-rounds.md
@@ -0,0 +1,70 @@
+# Microbiome
+
+Vincent B Young MD PhD, Ann Arbor
+
+Microbiome: community of microbes and environment they inhabit
+
+Microbiota: the microbes themselves
+
+Focus on C. Diff
+
+2-3% of healthy outpts have identifiable, toxin-producing C. Diff
+
+"Antibiotic Associated Colitis" 1977 papers that first described C. Diff related to abx, hospitalization, using hamsters as a model organism.
+
+"An Epidemic, Toxin Gene-Variant Strain of Clostridium difficile" 2005 NEJM
+
+C diff dx: PCR/LAMP, glutamate dehydrogenase testing (GDH) two vs three step, EIA for toxins.
+
+Controversy: Nucleic acid amplification tests (NAAT) cannot distinguish colonization vs infx (NAAT does detect toxin gene).
+
+20-30% of pts will test + for NAAT during hospitalization (?colonization, spore passing through).
+
+Controversy: should we use the most sensitive test (NAAT) to find even colonization, to control spread?
+Or use toxin tests up front, to catch the cases severe enough to produce detectable toxin? (i.e. use a purposefully less sensitive test that is possibly more specific for more severe dz)
+
+Classifying severe/complicated CDI:
+- Severe: WBC >15k, Cr >1.5x normal, absolute serum Cr >1.5 if no baseline available
+- Fulminant: hypotn, shock, ileus, toxi megacolon
+- Recurrent: 2-8wks from last positive specimen OR clinical response
+
+Studying microbiome:
+- Anatomy
+ - structure: "who is there?"
+- Physiology:
+ - actual function: "what is it doing?"
+ - potential function: "what can it do?"
+
+- 75-80% of tx cases do not recur; 20-25% of cases recur and have worse outcomes.
+ - theory: pts never return to normal microbiota, hence if restore normal microbiota -> cure
+
+
+- Hx fecal tx (Fecal Microbiota Transplantation, FMT)
+ - Pliny the Elder: fermented milk and fecal tx
+ - Ge Hong: 4th C
+ - another guy whose name I didn't catch: "yellow soup" == poop supernatant
+ - surgeons in 50s: successful fecal tx for abx-associated
+ - 2013 NEJM: 94% success rate for FMT (16pts), 30% vanco, trial ended early
+
+- Prior to FMT, community is "less diverse" than donors
+- FMT results in transfer of community structure to pts
+- Structure does NOT predict function - some pts who do *not* recover do have more diverse micriobiota, and some pts who *do* recover remain less diverse
+
+Microbiome -> metabolome, and metabolome significantly contributes to generation of spores vs inhibition of infx
+
+Mice != humans, mouse microbiome != human microbiome.
+Human feces known to be effective in tx CDI in humans is not effective in tx recurrent CDI in mice.
+Mouse FMT restores bile acid metabolism in mice, thought to be the main mxn.
+
+Jenna Wiens, PhD: ML for microbiome. 2018 Infx ctl and hosp epi, "A Generalizable, Data-Driven Approach to Predict Daily Risk of..."
+
+A generalizable approach vs a generalizable model.
+YES.
+(You can feed hospital-specific data to the same code, with some variation in preprocessing, and have a new model using a generalizable approach).
+
+Wiens now doing prospective work - YES again.
+
+
+Next steps:
+- moving from association to causation
+- precision medicine that includes host genome and microbiota genomes, etc.
diff --git a/learn/want-from-teachers.md b/learn/want-from-teachers.md
@@ -0,0 +1,35 @@
+# What This Millenial Wants From Instructors
+
+In a word: Context.
+
+I distrust textbooks (they're outdated before they're published),
+and wikis don't fare much better.
+Primary literature and review literature
+are wonderful, but have their own drawbacks.
+
+A brief word on each of these resources follows,
+then why I need an instructor to help me put it all together.
+
+Textbooks are meant to give context,
+either for an entire field
+or for a subfield.
+Some textbooks age well, while others do not.
+For myself as a trainee, the problem is that I do not know
+exactly which parts have aged well,
+and which should be updated.
+When I pick up a textbook myself,
+I hope the parts I read will
+be both accurate and up-to-date,
+but digging through the new primary literature
+to verify every piece would be a herculean task.
+
+UpToDate is a great wiki-style tool,
+with good tracing of provenance and upfront reporting of recency,
+and it comes in bite-sized chunks that are easy to use in a busy clinical context -
+but it also comes in bite-sized chunks,
+and in a busy clinical context clicking through the hyperlinked rabbit holes
+that would give me the rest of what I need to know is unlikely, and error-prone.
+Primary literature takes me right to the bottom of the rabbit hole,
+and as a physician-scientist,
+primary sources make me happy and comfortable -
+but it
diff --git a/projects/ai-for-mds-book.md b/projects/ai-for-mds-book.md
@@ -2,21 +2,42 @@
## Intro
-### Why
+### Why?
- It's happening
- Data availability
-
-### What is AI/ML
+---
+
+### What is AI/ML?
+
+
+AI
+
+Tools for doing an intelligent thing (we're not going to get into AGI: while very cool, it's not super practical, and tends to obfuscate the useful stuff --> lead to debates on the nature of humanity, etc.)
+
+Examples of simple AI: thermostats and cruise control.
+A human who would like to maintain a constant speed would
+A) pick the desired speed,
+B) note whether they were currently above or below that speed,
+and C) accelerate up or down to match the desired speed.
+This is an example of a simple rules-based system,
+a series of if-then statements (if above speed X, let off the gas. If below speed X, give it more gas).
+These are very useful, and you can get a lot of mileage out of them,
+but if decision-making is complex, then the series of interconnected rules become more complex,
+and increased complexity leads to brittle systems with limited functionality and a high maintenance burden.
+
+ML
+
+Tools for training a machine to do a thing by giving it examples from which it can learn. We will get into the major categories later, but in brief they are
-- AI: tools for doing an intelligent thing (we're not going to get into AGI: while very cool, it's not super practical, and tends to obfuscate the useful stuff --> lead to debates on the nature of humanity, etc.)
-- Examples of simple AI: thermostats and cruise control. A human who would like to maintain a constant speed would A) pick the desired speed, B) note whether they were currently above or below that speed, and C) accelerate up or down to match the desired speed. This is an example of a simple rules-based system, a series of if-then statements (if above speed X, let off the gas. If below speed X, give it more gas). These are very useful, and you can get a lot of mileage (har har) out of them, but if decision-making is complex, then the series of interconnected rules become more complex, and increased complexity leads to brittle systems with limited functionality and a high maintenance burden.
-- ML: tools for training a machine to do a thing by giving it examples from which it can learn. We will get into the major categories later, but in brief they are
- Classification: learn from example whether a thing is an X or a Y. E.g. is this a picture of a hotdog or a shoe? Based on labs, genes, and demographics, does this patient have disease X or disease Y?
- Regression: learn from example the quantitative value of a thing based on other data. E.g. based on local property values, square footage, amenities, for how much will this house sell? Based on lab values that are easy to obtain, can we predict another lab value that is difficult to obtain (e.g. bone marrow cellularity, blast percentage).
- Unsupervised: from a set of data, find clusters of similar data points. E.g. from a number of whole genome sequences, find patterns of similarity that can be further explored and correlated with phenotype. In an image, locate areas likely to be of interest.
+---
+
### Why you?
Data is king, queen, and jester
@@ -106,7 +127,7 @@ and that collecting that data would require hematopathologists,
and that hematopathologists are busy and expensive,
so they would build a system to make the task of labelling easy and fast.
As a part of this, they wanted to evaluate the quality of the labels they were going to obtain,
-and so performed a study to quantify agreement between and within hematopathologists for a select number of cellular entities and quantities.
+and so performed a study to quantify agreement between hematopathologists for a select number of cellular entities and quantities.
They found consensus in some things, when they were obvious,
but nonconsensus in non-obvious things.
@@ -117,8 +138,45 @@ This basic problem is repeated throughout machine learning in medicine,
from improperly or incompletely labeled radiographs
to chart-based diagnoses based on billing concerns more than clinical realities.
We need experienced clinicians to generate, evaluate, and mobilize data,
-and there is (and, I argue, will never be) a substitute for direct experience with real patients
+and there is not (and, I argue, will never be) a substitute for direct experience with real patients
(or, in the cases of pathology and radiology, the tissues and images of real patients, and consultation with the clinicians who ordered the studies).
+---
+
### How to learn
+There is an ever-growing number of learning resources available for the would-be data scientist.
+These range from in-person degrees at prestigious institutions,
+to online courses and certifications,
+to books,
+blog posts,
+and code examples that themselves range from the
+line-by-line hands-on approach
+to the highly cerebral and theoretic.
+
+Which resources you choose to use will depend largely on your goals.
+
+If you want to be a theoretician,
+you may start from the ground up
+and study the core mathematics and computer science.
+
+If you want to be an engineer,
+that is, one who applies the tools to real world problems,
+you may prefer a hands-on approach at first,
+and dip into the theory as needed to troubleshoot deeper issues.
+
+You may want to be neither of these,
+but instead your goal is to be a responsible citizen
+within your medical field,
+with enough understanding to critique the
+academic articles and,
+perhaps more importantly,
+software tools employed within your practice and institution.
+This may be analogous to the way and reasons we learn
+statistics in medical school,
+not to become statisticians,
+but to know how to speak with statisticians,
+how to collaborate,
+how to understand what we read,
+and spot poor practice
+or misleading claims.
