AI-Assisted Data Extraction with Large Language Models for Systematic Review and Meta-Analysis

Introducing AI-Assisted Data Extraction, a free, open-source solution for human-in-the-loop data extraction with large language models. Stop wasting time and extract data fast with AIDE!

Please Like and Subscribe to support the channel! @LearnMetaAnalysis

0:00 Intro
0:29 Installation
1:03 LLM Set-up
1:58 Analysis
2:50 Results
4:28 Time savings
5:17 Coding Form
5:58 Study 2
6:55 Larger example
7:49 Summary

[zachariahmarrero]

This is a great start but what you need to be doing is also measuring the detection rate for different prompts. For example, content about implications for theory might exist in several places and the LLM may fail to reference many of them. That would be false negatives. Conversely, it can make reference to irrelevant information, false positives. Measuring this would probably be best approached by applying the method to lots of research papers when the prompt is the same. Ideally, you also do this for 100s-1000s of prompts and then train a secondary model to predict the estimated detection rate directly from the prompt. then later, you can highlight or warn the user when a particular prompt is estimated to result in higher failure rate. That’s going to be really important because most people are still not thinking about “AI” in these terms.

[CyclonerAI]

Very cool! Have you looked into Automated Reasoning yet for contracts and documentation? It looks like AWS still had not released it even though they claim they have it available on the US West Oregon region. Looks like amazing stuff!

[avocado]

Thank you for your hard work. We would appreciate it if the software could be updated to allow users to utilize a remote Ollama server using an API key, base URL, and Model Name. Once again, thank you.

[JTB-zw]

Hey, I'm experiencing the same issue with the reply “No Response found for the Prompt” in the later prompts of my coding form. I haven’t found a proper fix so far. Prompts vary in complexity (about 29 prompts in the codingform). Previous responses often get cut off as well. In “R”, the generated response always shows a maximum output token limit of 8.192, regardless of the Gemini model I use. However, Google states very different values for maximum output tokens online (often over 65.000). Is there any way to change this, or is it ultimately a limitation of Gemini? Do you have any other suggestions for this issue? Thank you very much in advance for your help!

[carloszepeda]

Great work on AIDE—congrats! Quick question: when I click “Analyze PDF” after uploading my Excel, I get “API Error: HTTP 400 Bad Request.” Any idea why?

[aymannadeem]

Which Gemini model do you suggest, I am getting a reply of "No Response for the prompt" in the latter prompts of my coding form. The issue goes away often when I just click Analyze. But I have to sometimes change models to surpass this error. My coding form has 40-60 prompts (1/4 coding forms ) and the model I am using is Gemini 2.5 Flash Preview 05-20, I have also tried Gemini 2.0 Flash.

[EmiliaMadueme-fr]

Hi. After I click "analyze PDF" on AIDE, I got the error message: API Error: HTTP 400 Bad Request. I am not quite sure how to resolve this. It's my first time doing a meta analysis research.

[vytauskas]

I'm having an issue with this. I've followed all the steps precisely as you've shown, but I continuously get a "API Error: HTTP 400 Bad Request." message that appears in the prompt windows whenever I attempt to analyse a PDF. I've tried with API from both Mistral (which didnt even load any models) as well as Google/Gemini and I get the same result. Could you help me do a bit of troubleshooting for this? Thank you :)

[krishnasaketh]

Hey, I've been getting API Error: HTTP 400 Bad Request error quite often these days. I intially was able to fix it by removing the formatting in my coding form but that trick isnt working anymore, any suggestions ?

[manu]

Really cool project, but can't get it to fetch a Gemini model. I made sure there's no space at the end