Problem reading accurate motorbike tilt angle

Hello,

I try to reading motorbike tilt angle and I experiment a problem I also have with MPU-6050 in fact.
These chip seem to work well on desktop but using it on vehicle seem to be a challenge.
As ICM-20948 seems to be usable on drone, I though it was dedicated to give accurate tilt angle but it seems to be more complicated than it appears.

First, I tested it on my desktop, applying the same movements to my Samsung S10+ (embedding a LSM6DSO) and the ICM-20948 breakout.
The result is this one:
as you can see, the waveforms perfectly match.

Here, I did the same mounting both devices on my motorbike.
Between 547 and 807, I really was in stable turn, with the same tilt angle as given by my S10+. But the ICM-20948 was seriously drifting to 0 and the offset accumulated at 807 seems to appear at 911.
Between 1639 and 1925, I was accelerating as the turn decreases. S10+ gives an understandable value but the ICM-20948 dramatically drift to 0, probably while the z accel increases.

I guess the problem could come from the pitch drift correction made with the g acceleration. And as on motorbike, as well as on drone, a turn cause the g acceleration to increase, maybe this algorithm causes the pitch reading to be false.

I hope I was clear enough, do you have an idea about the problem and how to solve it?