Health probe marks instances as unhealthy but the aren't

I use a VM scale set for my node application. My app has an action which is public accessible via www.mydomain.com/api/healthcheck and prints just some json. When I configure my health probe to use TCP protocol, everything works fine and also my api returns me the expected json (and status 200). However, when I now switch my health probe to use HTTP and path=/api/healthcheck, my website isn't accesible anymore (ERR_CONNECTION_TIMED_OUT... I guess the loadbalancer takes out all instances because the health probe tells him that every instance is unhealthy)

I use nginx in front of my node app, but I also tried (for testing) to configure my LoadBalancer to route port 80 to backendport 8080 (where my node app is running on every machine, so I can avoid nginx proxy). But I get the same behaviour.

I'm out of ideas why my custom health check doesn't work. Hope you can help.

Edit: For testing, I did the following:

• run another nodejs app on port 3000 on every VM, which just prints "hello world" (without nginx proxy!)
• create a LB rule for port 3000 and also configure my NSG to allow :3000 for all
• at the beginning, my health probe is configured to use tcp
• result: mydoamin.com:3000/hello is available (prints hello and returns 200)
• now I configure my health probe to use http-protocol, port 3000 and location /hello.
• result: my whole web app isn't available anymore