@Yveaux thanks for your answer.

I used the sketch on the Mysensors'si7021 page, so the vcc level is read at the end of the loop, before going to sleep and after sending the temp/hum values.
The only modification I made are setting up the voltage values.
As my 3.7V Li-Ion battery is at 4.2V when fully charged and at 3V when almost completed discharged, this how I configured it in the sketch:

#ifdef REPORT_BATTERY_LEVEL
#include <Vcc.h>
static uint8_t oldBatteryPcnt = 200;  // Initialize to 200 to assure first time value will be sent.
const float VccMin        = 3.0;      // Minimum expected Vcc level, in Volts: Brownout at 1.8V    -> 0%
const float VccMax        = 4.2;  // Maximum expected Vcc level, in Volts: 2xAA fresh Alkaline -> 100%
const float VccCorrection = 1.0;      // Measured Vcc by multimeter divided by reported Vcc
static Vcc vcc(VccCorrection); 
#endif

I didn't apply any correction because read and reported VCC was very close, and I'm just testing this for now.

I didn't removed the onboard regulator nor the power led for now. Isn't the regulator only used when applying voltage to the RAW pin?

@mfalkvidd I thought of this too, but I opened the windows this morning, 11°C here this morning, and the battery level stayed around 50%.

Here are some graphs showing it's not temperature related:

Even with the 0.1uF capacitor, I had wrong values.
So, I've gone with the voltage divider method, and it seems more reliable, at least for my case.
4.7V max to VCC, R1 1M, R2 330k, another 0.1uF capacitor in parallel with R2.
I don't need a very good accuracy, just a trend in order to know when to change my batteries.
Thanks for your help.