Big data

# Big data
## the inevitable problem with overdiagnosis
### Claus Thorn Ekstrøm<br>UCPH Biostatistics<br>.small[<a href="mailto:ekstrom@sund.ku.dk">ekstrom@sund.ku.dk</a> ]
### August 21st, 2018<br>.small[Slides @ <a href="www.biostatistics.dk/talks/">biostatistics.dk/talks/</a>]

---

---

---

# "With enough data, the numbers speak for themselves."

---

---

# The lure of big data

*   Absurdly accurate results
*   Everything can be captured
*   The correlations will tell all - causality less important
*   Statistical models not needed
]

*   Cheap to collect
*   Likely to influence multiple diseases
*   Stable
]

---

---

# How are big data currently used?

1.  Larger samples - better coverage
2.  More precise diagnoses
3.  Earlier scans
4.  Scan for multiple features/diseases/abnormalities
5.  Lifetime trajectories
7.  Redefine disease definitions

---

# Large samples

Better coverage. `$N=$` all? `$P$` massive

![](preventing-overdiagnosis-2018_files/figure-html/unnamed-chunk-1-1.png)

---

# More precise diagnoses

*"If screening positive do I have the disease?"*

Positive predictive value:

`$$\begin{split}PPV &= P(\text{disease} | \text{screen positive}) \\
&= \frac{P(\text{screen positive}|\text{disease}) \cdot  P(\text{disease})}{P(\text{screen positive})} \\
&= \frac{\text{sensitivity} \cdot  P(\text{disease})}{\text{sens.} \cdot P(\text{disease}) + (1-\text{spec.}) \cdot P(\text{no disease})}
\end{split}$$`

---

# Honest PPV

*"If screening positive do I have a disease that will give me problems?"*

`$$\begin{split} 
hPPV &= P(\text{problematic disease} | \text{screen pos.}) \\ & = P(\text{prob. dis.} | \text{scr pos., disease})\cdot P(\text{dis.}|\text{scr pos.}) \\ & + P(\text{prob. dis.} | \text{scr pos., no dis.})\cdot P(\text{no dis.}|\text{scr pos.})
\end{split}$$`

---

# Honest PPV

*"If screening positive do I have a disease that will give me problems?"*

`$$\begin{split} 
hPPV &= P(\text{problematic disease} | \text{screen pos.}) \\ & = \underbrace{P(\text{prob. dis.} | \text{scr pos., disease})}_{(1-OD)}\cdot \underbrace{P(\text{dis.}|\text{scr pos.})}_{PPV} \\ & + \underbrace{P(\text{prob. dis.} | \text{scr pos., no dis.})}_{0}\cdot \underbrace{P(\text{no dis.}|\text{scr pos.})}_{} \\
&= (1-OD)\cdot PPV
\end{split}$$`

---

---

# Multiple features

![](preventing-overdiagnosis-2018_files/figure-html/unnamed-chunk-2-1.png)

---

# Lifetime trajectories

<div id="htmlwidget-6274bf6fc4f8bfad828a" style="width:960px;height:400px;" class="grViz html-widget"></div>
<script type="application/json" data-for="htmlwidget-6274bf6fc4f8bfad828a">{"x":{"diagram":"\ndigraph G { \n    bgcolor = \"transparent\"\n    node [shape = circle, fontcolor = \"white\", color = \"white\"]\n    edge [color = \"white\"]\n\n  rankdir = LR;\n  subgraph {\n    A -> C -> D -> B\n    G -> E\n    X -> Y\n    // note that rank is used in the subgraph\n    {rank = same; A; X;}\n    {rank = same; B; Y;}\n    {rank = same; C; E;}\n  } /* closing subgraph */\n}\n","config":{"engine":"dot","options":null}},"evals":[],"jsHooks":[]}</script>

---

# Modeling over-diagnosis

---

# Step 1: Redefine the outcome(s)

When do you *really* have the disease?

Diagnosis `$\leadsto$` Treatment `$\leadsto$` No problem

Diagnosis `$\leadsto$` No treatment `$\leadsto$` No problem

What is *"no problem"*?

No amount of big data + machine learning techniques will fix this!

---

# .black[Step 2: Define a loss function]

---

# Step 3: Timing

*When* do we have a diagnosis? Multiple outcomes ...

<div id="htmlwidget-bd56604ec998b788a88b" style="width:960px;height:130px;" class="grViz html-widget"></div>
<script type="application/json" data-for="htmlwidget-bd56604ec998b788a88b">{"x":{"diagram":"\ndigraph G { \n    bgcolor = \"transparent\"\n    node [shape = circle, fontcolor = \"white\", color = \"white\"]\n    edge [color = \"white\"]\n\n  rankdir = LR;\n  subgraph {\n    A -> C -> D -> B\n    // note that rank is used in the subgraph\n    {rank = same; A; X;}\n  } /* closing subgraph */\n}\n","config":{"engine":"dot","options":null}},"evals":[],"jsHooks":[]}</script>

Competing risks

---

# Step 3: Timing

*When* do we have a diagnosis? Multiple outcomes ...

<div id="htmlwidget-27980fe59946d09e419e" style="width:960px;height:130px;" class="grViz html-widget"></div>
<script type="application/json" data-for="htmlwidget-27980fe59946d09e419e">{"x":{"diagram":"\ndigraph G { \n    bgcolor = \"transparent\"\n    node [shape = circle, fontcolor = \"white\", color = \"white\"]\n    edge [color = \"white\"]\n\n  rankdir = LR;\n  subgraph {\n    A -> C -> D -> B\n    // note that rank is used in the subgraph\n    {rank = same; A; X;}\n  } /* closing subgraph */\n}\n","config":{"engine":"dot","options":null}},"evals":[],"jsHooks":[]}</script>

Competing risks ... of overdiagnosis

*If I am highly likely to die from XXX in the next 3 weeks should I worry about YYY?*

---

background-image: url(pics/ghostbusters.jpg)
background-size: 100%
class: middle, center

# The beams MUST cross