digipres.qmd

---
title: "Empowering Medical Professionals" 
subtitle: "Leveraging GPT Models for Informatics Solutions"
author: "Christopher Snyder, MD, PhD"
date: "2024-10-02"
slide-number: true
format:
  revealjs:
    scrollable: true
    theme: solarized
    preview-links: true
    css: https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.1.1/css/all.min.css
    include-in-header:
      - text: |
          <link href="https://cdn.jsdelivr.net/npm/@jupyter-widgets/html-manager@1.0.6/dist/embed-amd.css" rel="stylesheet">
          <script src="https://cdn.jsdelivr.net/npm/@jupyter-widgets/html-manager@1.0.6/dist/embed-amd.js"></script>
jupyter: python3
execute:
  freeze: auto
  enabled: true
  cache: false
---

<style>
.blue {
    color: blue;
}
</style>

<style>
.yellow {
  color: yellow;
}
</style>

<!-- stock themes
[default, beige, blood, dark, league, moon, night, serif, simple, sky, solarized] -->

<!-- # ipynb-shell-interactivity: all -->

## Financial Diclosures

I have no financial disclosures and no conflicts of interest to declare.


## AI Hootenanny

::::: columns
::: {.column width="50%"}
### Pervasive Hype 
- focus is on heralding future innovation
- my Natural Language Processing(NLP) professor around 2018 lamented that his work seems to follow him home in the news
- will reshape our daily lives - similar to industrial revolution - eventually reaches "Singularity" - potential to "transform healthcare"
:::

::: {.column width="50%"}
### Sparse Real-World Problem Solving
- "in the wild" either doesn't happen or isn't talked about
- [**Can anyone think of any examples where AI could *maybe* be applied?**]{ .blue}
:::
:::::




## Like Most Prophecies, 'AI Forcasts' are Hardly Actionable

-   What should I do?
-   When should I do it?

::: {.notes}
  what constitutes an adaptive response
  depends on your assesment of Two Variables
:::

## A Timeline for AI

::: {.r-stack}
![](src/intro/ai-hype01_boombust.jpg){.fragment .fade-in-then-out}

![](src/intro/ai-hype02_nextbigthing.jpg){.fragment .fade-in-then-out}

![](src/intro/ss-punnet-fragile0.png){.fragment .fade-in-then-out}

![](src/intro/ss-punnet-fragile1.png){.fragment .fade-in-then-out}

![](src/intro/ai-hype03_colorfultimeline.png){.fragment .fade-in-then-out}
:::



## Objectives

1. Formative experience applying AI to pathology call
2. Why AI will let anyone do informatics
3. Demos applying AI to problems from real-world resident tenure
4. Classification System for AI application 
5. A few suggestions about the future

<!-- 
### outline
1.  Outline the mechanisms by which AI adapts and accelerates real world workflows *today* (from my tenor as a pathology resident)
2.  Explain conceptually: how to anticipate when AI will be useful to apply to a particualr circumstance.
3.  Demonstrate "hands-off" informatics implementation of AI solutions, also made possible by AI.
4.  Give a Theoretical Basis: why proficiency in ai-tools (in-principle) should be (a) empowering and (b) eventually universal, vis-a-vis literacy
5.  Introduce Diagramatically a classification system for AI-Application facilitating computer-tasks.
6.  Automate Call Log Entry via an AI Application. <!-- - entry of call metadata into the pathology resident call-log form from simulated call audio. -->
::: { .smaller .indent=2em}
-   if time, show how an AI agent can be used to append, modify, save, query, analyze past log-entries.
-   if time, show easy integration with downstream informatic-endpoints e.g. email notification for impending sample expiration, ai entry summarization and report generation. -->


## Calibrating Perceptions of Informatics: *a prerequisite to discussing change*

::::: columns
::: {.column width="50%"}
#### What is Informatics to *You*

-   baseline *status quo ante* AI
-   perceptions vary 
:::

::: {.column width="50%"}
<img src="src/intro/aram-mojtabai-hacking.gif" alt="Hacking GIF" style="width:100%;max-width:100%;"/><p style="margin-top:10px;font-weight:bold;">
honest impressions, 2012
</p>
:::

:::::


::: notes
Truth -No one ever gets dramatically more efficient than a beginner Truth- No one is naturally advantaged or inclined. Informatics is naturally inclusive
:::


## Informatics is Hard for Everyone
Always has been 
<!-- , \[~~Always will be~~\] -->

<!-- ![](src/intro/xkcd_terminal.webp){width="50%" max-width="50%"} -->

#### **the terminal**
-   intimidating, unforgiving, powerful

##### Previous approach
```{bash}
tar -czvf file.tar.gz mydata.csv 
```


##### AI aided approach
```{bash}
gh copilot suggest "please compress my file with `tar` command
  Use the most common, generic approach.
  You know what I want...Just do the thing, please."
```




## Gene Panels, Spot the Difference

![](src/gene-panels-03.png){width="30%"}

![](src/gene-panels-02.png){.r-stretch}


## Gene Panels, Spot the Difference

::::: columns
::: {.column width="25%"}
[chatgpt thread](https://chatgpt.com/share/66f2e625-ab44-8005-a883-686a92017a51)

:::

::: {.column width="75%"}
![](src/gene-panels-01.png)
:::
:::::



## "Just Autocomplete"

#### [A Slippery Slope]{style="color: blue;"}

|  |  |
|----|----|
| **Text** Autocomplete | -> next word |
| **Language** Translation | <X> in Spanish would be -> |
| **Information** Transformation | \- data representation \<-\> (img, txt, code.py) |
|  | -> board material -> podcast discussing it |
| **Instruction** Implementation | goal -> code achieving goal |

# WUSTL Weekly Biogram Report Analysis

data from <https://diorama.wustl.edu/weekly-report>

## Weekly WUSTL Biogram

![](src/biogram-01.png)

## LLM-Driven Data Slicing (1/3)

![](src/biogram-alldata.png)

## LLM-Driven Slicing (2/3)

![](src/biogram-justviruses.png)

## LLM-Driven Slicing (3/3)

![](src/biogram-gettingworse.png)

## LLM-Driven Data (?)Dicing

![](src/biogram-piegraph.png)

# Intermission

<!-- ## Progress Outline

-   [x] Introduction to "Computers that Know What You Mean"
-   [ ] Why would anyone want to "talk to computers"?
-   [ ] Overview of current tools and methologies
-   [ ] Example walkthrough of a simple informatics task
-   [ ] Recommendations for leveraging this and future transformative technologies -->

# Why Should I Talk to Computers? How is that empowering?

## Communication Critical for Humans

::::: columns
::: {.column width="50%"}
-   Children Universally Geared toward learning language
-   Why?
    -   tool for learning
    -   exploring feasibility
    -   coordinating resources
    -   accomplishing goals
:::

::: {.column width="50%"}
![Charlie \~ 6 weeks](src/language/charlie01.png)
:::
:::::

## Communication Critical in the Workplace

::::: columns
::: {.column width="50%"}
-   Doctors - Patient Bedside <!-- - Describing symptoms --> <!-- - Understanding treatment options -->
-   Multi-Diciplinary Care Teams <!-- - Discussing prognosis --> <!-- - Coordinating care -->
-   Blood Bank <!-- - Requires >1 person -->
:::

::: {.column width="50%"}
<!-- ![](src/language/meeting.webp) -->

![](src/language/meeting_skinny.png)
:::
:::::

## Communication is Empowering

| 
| 

<!-- |--------------------|------------------------| -->

| Benefit | Communication Example |
|----|----|
| Economies of Scope | \-  enlisting adults<br>- Multi-disciplinary <br>[ - **learning from data**]{.fragment fragment-index="1"} |
| Economies of Scale | \- cooperation <br>- minimum staffing, redundancy <br>[ - **automating tasks**]{.fragment fragment-index="1"} |

## AI: Key to Efficiency

-  By lowering the barrier working with computers, and by expanding the pool of practitioners who can do so...
-  AI most likely **will remove essentially all present operational inefficiencies** of today
-  everything is worth the squeeze when you have a juicer






## AI-Informatic Approaches

0.  task to code (by hand)

1.  task as prompt

2.  task to code

3.  code as prompt

![](src/tools/aitools-flowchart-byhand.png){width="100%"}

## AI-Informatic Approaches
0.  task to code (by hand)
  -   tedious, "too literal": always requires 100% specification
1.  task as prompt
  -   easy, difficult to specify what you mean precisely enough
2.  task to code
  -   Can jumpstart coding, modifying easier than writing
3.  code as prompt
  -   sweet spot: clearly specifies what you want the output to "look like" while leaving the details to the model



## UDS Parse {.scrollable}

####  Task: parse UDS screen results for positive patient+drugs resulted

<!-- [uds_notebook](http://localhost:8888/doc/tree/drugscreen_result_parsing.ipynb){preview-links=true} -->

[uds_notebook](file:///Users/christophersnyder/Projects/digital-path-talk-2024/uds_parse/drugscreen_result_parsing.html){preview-links="true"}

SCREEN_RESULT_COMMENT

"The following compounds were detected:

Tetrahydrocannabinol (THC)

Repeated and verified. Medical Director review to follow." "The following compounds were detected:

Tetrahydrocannabinol (THC)

Repeated and verified. Medical Director review to follow."

"The following compounds were detected:\
Benzoylecgonine (BEG) Fentanyl Tramadol O-Desmethyl tramadol Repeated and verified. Medical Director review to follow."

"The following compounds were detected:

Amphetamine Methamphetamine Alprazolam (Xanax) Lorazepam-glucuronide

Repeated and verified. Medical Director review to follow."

"The following compounds were detected:

Morphine Morphine-3-glucuronide Fentanyl

Repeated and verified. Medical Director review to follow." "The following compounds were detected:

Tetrahydrocannabinol (THC),

Repeated and verified. Medical Director review to follow." "The following compounds were detected:

Oxycodone,

Repeated and verified. Medical Director review to follow."

"The following compounds were detected:

Amphetamine

Methamphetamine . . .

## 1. Task as Prompt

::::: columns
::: {.column width="50%"}
-   try [chatgpt](https://chatgpt.com/share/66faa53f-91fc-8005-b9a6-243465db3d1d) directly
-   clearly feasible
-   difficult to specify what I mean precisely enough
:::

::: {.column width="50%"}
![caption: screenshot of thread with chatgpt](src/tools/uds-chatgpt-01.png)
:::
:::::

## 0. Task to Code (by hand)

<!-- ```{python} -->

#### code to parse UDS screen results for positive patient+drugs

``` {.python code-line-numbers="3-5|11|1-100"}
def parse_section(section: str) -> list[str]:
    """Parses a section of the screen result for detected compounds."""
    start = section.find("detected:")
    end = section.rfind("Repeated")
    lines = section[start + len("detected:") : end].strip().split("\n")
    compounds = [l.strip() for l in lines if l.strip()]
    return compounds

results_file = Path( "screen_result.txt" )
result_text = results_file.read_text()
sections = result_text.split('"')[1::2]
parsed_sections = [parse_section(section) for section in sections]
print(*parsed_sections, sep='\n--------------------\n')
```

#### !"overly specific", hyper granular, brittle

``` {.text .scrollable}
['Tetrahydrocannabinol (THC)']
--------------------
['Tetrahydrocannabinol (THC)']
--------------------
['Benzoylecgonine (BEG)', 'Fentanyl', 'Tramadol', 'O-Desmethyl tramadol']
--------------------
['Amphetamine', 'Methamphetamine', 'Alprazolam (Xanax)', 'Lorazepam-glucuronide']
--------------------
['Morphine', 'Morphine-3-glucuronide', 'Fentanyl']
--------------------
['Tetrahydrocannabinol (THC),']
--------------------
['Oxycodone,']
--------------------
['Amphetamine', 'Methamphetamine']
--------------------
['Amphetamine', 'Methamphetamine']
--------------------
['Cocaine', 'Benzoylecgonine (BEG)', 'Fentanyl', 'Ketamine']
--------------------
['Nordiazepam,']
--------------------
['Oxycodone', 'Lorazepam-glucuronide', 'Tetrahydrocannabinol (THC)', 'Fentanyl']
--------------------
['Methadone', 'Fentanyl']
--------------------
['Tetrahydrocannabinol (THC)', 'Fentanyl']
--------------------
['Tetrahydrocannabinol (THC)']
--------------------
['Hydrocodone']
--------------------
['Oxycodone']
--------------------
['Methadone,', 'EDDP (Methadone Metabolite),', 'Fentanyl,']
--------------------
['Tetrahydrocannabinol (THC),']
--------------------
['Amphetamine,']
--------------------
['Benzoylecgonine (BEG)', 'Tetrahydrocannabinol (THC)']
--------------------
['Methadone', 'EDDP (Methadone Metabolite)', 'Fentanyl']
--------------------
['Tetrahydrocannabinol (THC)', 'Oxycodone', 'Hydrocodone']
--------------------
['Methadone', 'EDDP (Methadone Metabolite)', 'Bupropion', 'Hydroxy-bupropion']
--------------------
['Tetrahydrocannabinol (THC)']
--------------------
['owing compounds were detected', 'Tetrahydrocannabinol (THC)']
--------------------
['7-aminoclonazepam']
--------------------
['Amphetamine,', 'Methamphetamine,', 'Tetrahydrocannabinol (THC),', 'Oxycodone']
--------------------
['Methamphetamine']
--------------------
['Amphetamine,', 'Methamphetamine,', 'Morphine,', 'Morphine-3-glucuronide']
--------------------
['Amphetamine']
--------------------
['Venlafaxine(Effexor)']
--------------------
['Cocaine,', 'Benzoylecgonine (BEG),', 'Oxycodone,', 'Fentanyl']
--------------------
['Benzoylecgonine (BEG),', 'Tetrahydrocannabinol (THC),', 'Methadone,', 'EDDP (Methadone Metabolite),', 'Fentanyl']
--------------------
['Amphetamine,', 'Methamphetamine']
--------------------
['Amphetamine,', 'Methamphetamine']
--------------------
['Lorazepam-glucuronide.']
--------------------
['Benzoylecgonine (BEG)']
--------------------
['Lorazepam-glucuronide.']
--------------------
['Methylphenidate (Ritalin)']
--------------------
['Lorazepam-glucuronide.']
--------------------
['Clonidine']
--------------------
['Tetrahydrocannabinol (THC)', 'Norbuprenorphine']
--------------------
['7-aminoclonazepam']
```

## 2. Task to Code

uds_parse/drugscreen_result_parsing.ipynb

![](src/tools/uds-chatgpt-code-02b.png)

[chatgpt uds task -> code](https://chatgpt.com/c/66faa511-42c4-8005-86c0-ae32a866b1b3){preview-link="true" style="text-align: center"}


## 3. Code as Prompt

![Coding without Rigid Rules](src/tools/uds-marvin-03.png)




# Suggestions

## "low hanging fruit" for AI application

- Flow lab dashboard
  - make "pull" processes "push" processes


- Txfsn Service AI Assistant for inpatient
  - logging, updating, retrieving, summarizing log info
  - "Jarvis-like" interface 
  - addtl functions:
    - notifications
    - auto rounds printouts


<!-- ## Pursuit Angle -->

## Scales of Leadership {.smaller}

::::: columns
::: {.column width="33%"}
### Tactical
-   update specific lectures
-   ai elective
:::

::: {.column width="33%"}
### Operational
-   gpt-api part of onboarding
-   safe environment for prototyping and learning
-   incentivize process improvement
:::

::: {.column width="33%"}
### Grand-Strategic
-  what mentality or philosophy is most adaptive in the long-term?
<!-- -   How to cope with change that is accelerating? -->
<!-- -   What short term targets keep us relevant in the long-term? -->

:::
:::::

## trivia call-back

<!-- ["Beast Mode"](src/pusuit_angle/BeastQuake720p.mp4) -->

<video width="100%" height="auto" controls>

<source src="src/pusuit_angle/BeastQuake720p.mp4" type="video/mp4">

Your browser does not support the video tag. </video>

## 

![](src/pusuit_angle/ss_before.png)

## 

![](src/pusuit_angle/ss_after.png)

## 

![](src/pusuit_angle/ss_directly-chasing-the-ball.png)

## 

![](src/pusuit_angle/ss_slower-adopters-must-plan-farther-ahead.png)


## Gist

* We should prepare for the future like it’s already here. And, by the time we are ready, it will be.

* It is exactly because medicine will have the most obstacles to AI adoption that we need to plan the farthest ahead.

* Each *future* barrier to adoption makes less relevant the *present* technical limitations of AI


## on free time

#### The rate of return on an investment of your time is perhaps the highest ever

![](src/conclusion/xkcd_11th_grade.png)


## Final Remarks
|
|
### AI - The Peripheral Nervous System of the Laboratory

- AI removes friction in informed decision making
  - ease of information content appreciation
  - more contextualized and timely process adaptation
- decisions only as informed as access to information
- most data not perceptible to the 5 senses


## Thank you!