Metrics are a "point in time" measurement. The number you're seeing is true for an instant - when the measurement is taken.
My guess here (and it is a guess) is that there are times when one of your instances has disconnected (there will be timeouts, reboots, kernel updates, many things that can cause a brief disconnect) and during that time there will be one less connection to EFS; and that may be when the metric is calculated.
Similarly, an unexpected drop in a connection and a subsequent reconnection will see a situation where EFS thinks the old connection is still alive (it will timeout eventually) as well as the new connection. In that case you'll see four connections for a short period.
However: It's impossible to tell without detailed logs from the operating system and similarly detailed logs from EFS. Our support team can assist with troubleshooting the EFS side so I'd recommend that you raise a support case with them to dive deeper.
From a reliability and performance perspective I would enable detailed filesystem and network logging on your instances (temporarily) to try and determine if and why connections are dropping.
Send efs backup to S3Accepted Answerasked 10 months ago
How to avoid circular dependencies between EFS and access point in file system policy of Cloudformation templateasked 17 days ago
EFS throughput mounted on many instancesAccepted Answerasked 24 days ago
Is it possible to make EFS publicly accessible?
EFS high metadata IOasked 4 months ago
[Announcement] Amazon Elastic File System (Amazon EFS) is now available in the AWS Asia Pacific (Jakarta) Regionasked 8 months ago
EFS File system policy with IAM Instance Profile collisionAccepted Answer
EFS IA timerasked 3 years ago
EFS Throughput not exceeding 1MBasked 8 months ago
EFS Client Connections Metric Fluctuationsasked 2 months ago