Coach DrépanoCoach DrépanoCrisis risk algorithm
Trévor · version 2.251.85

Crisis risk algorithm

All of Trévor’s algorithms: form drop, today’s form, night and oxygen, blood, treatments →

Every morning, Trévor estimates the probability of a crisis within 48 hours. This page describes each step of the calculation, the values used and what changes after a transfusion.

The calculation, end to end

Four families of data enter the calculation. Each measurement is compared with your own normal, then weighted. The sum becomes a probability.

Watch or ringheart, oxygen, sleep, steps What you reportpain, crises, symptoms Your recordtransfusion, hemoglobin, treatments Weather and aircold, heat, particles 1. Gap from YOUR normalmedian and median deviation over the last 30 days,3-day trend; at least 7 days of measurements 2. One weight per signalstarting model (literature), cohort,personal model from 4 crises and 60 days 3. Sum on the model scalebase −2.2 + sum (weight × signal); weights × 1.6,except protections (hydroxyurea, transfusion) 4. Probability1 / (1 + e^−sum), never above 99% A transfusionstill holding? yes After a transfusionearlier signals reduced, protection added,risk capped (see the next flowchart)unless fever, breathing difficulty, chest pain no Score, level, explanationslow below 30 · moderate 30 to 59 · high from 60the 4 signals that weigh most are shown
The score is calculated on your phone with the same weights as on the server. The app reloads the model every 5 minutes while the screen is open.

After a transfusion

A transfusion or an exchange lowers the share of hemoglobin S in the blood. The calculation turns this fact into three effects, all governed by how long the transfusion holds.

The holding time is 5 days per unit : 35 days for 7 units, 40 days for 8. It is limited to between 14 and 56 days, and equals 28 days when the number of units is unknown.

Last transfusion recordedprofile, history or emergency card; without a transfusion,nothing below applies Days elapsed d, number of units nholding time T = 5 days × n, between 14 and 56 daysT = 28 days without a number of units Earlier nightsreduced by60 % × e^(−d/6)night oxygen,crisis and painof the last 7 daysup to 21 days Protection−3 × e^(−d/6)−2 × (1 − (d/T)²)strong on day one,almost stable afterwards,zero after T days Fading effectpenalty 1 − e^(−d/T)full in a program,half otherwise, andonly during90 days Model probabilitysum, weights × 1.6, sigmoid (see the previous flowchart) Fever, breathing difficultyor chest painreported? no Cap = 99% − 95% × (1 − (d/T)²)4% on day one, rises back toward 99% at the end of T yesno cap Score = the smaller valuebetween the model probability and the cap
The cap is deliberately strict: the displayed risk never exceeds what the transfusion's protection allows, unless a warning sign is reported. An alert must never be smothered.

Risk cap by day

Risk cap0 %25 %50 %75 %100 %071421283542days since the transfusion2 units7 units8 units
For 7 units (red), the cap stays below 20% for two weeks. For 2 units, protection stops at 14 days. Values are percentages of the 48-hour crisis risk.
Day2 units7 units8 units
0 d4 %4 %4 %
1 d4 %4 %4 %
3 d8 %5 %5 %
7 d28 %8 %7 %
14 d99 %19 %16 %
21 d99 %38 %30 %
28 d99 %65 %51 %
35 d99 %99 %77 %
40 d99 %99 %99 %

Transfusion program or not

In a programtransfusion program? yes no In a programtransfusion delay counted (weight 0.8)full fading effect, over T daysprotection and cap as above Not in a programno transfusion delayhalf the fading effect, 90 daysprotection and cap as above
You are in a program if you set it up, or if your history shows it: at least 3 transfusions in 13 months, on average at most 10 weeks apart.

The main weights

Each signal weighs on the model scale. A positive weight raises the risk, a negative weight lowers it. The weights shown are those of the starting model, before the 1.6 factor (except for protections). The personal model then adjusts them from your data.

SignalWeight
Recent transfusion−3,0
Transfusion still holding−2,0
Fading effect of the transfusion+1,4
Reported breathing difficulty+1,0
Recently reported pain+0,9
Rising skin temperature+0,9
Oxygen below your threshold (92% by default)+0,8
Reported fever+0,8
Transfusion late on your program+0,8
Night oxygen dip+0,7
Oxygen higher than your normal−0,6
Crisis in the last 7 days+0,6
Hydroxyurea taken regularly−0,55
Higher resting heart rate+0,55
Low hemoglobin+0,5
Days of pain in the week+0,45
Heart rate variability higher than your normal−0,45

The model counts about fifty signals: sleep, steps, measured stress, breathing, weather, air quality and reported symptoms complete this list.

The personal model

From 60 days of follow-up and 4 crisis days, Trévor trains a model of your own. It starts from the starting weights and departs from them only when your data justify it. It is kept only if its ability to tell crisis days apart exceeds 0.65 (AUROC, validated on blocks of time). Real example: over 174 days of follow-up and 6 crises, it reached 0.94.

The personal model, the cohort model and the starting model are combined according to their reliability. The starting model always keeps a weight.

What this algorithm cannot do

Weights that are not validated. The starting weights come from the literature on watches and sickle cell disease. The cap after a transfusion, the duration of 5 days per unit and the reduction of earlier signals were set from one patient's experience, without a study to confirm them.
A low risk does not rule out a crisis. While a transfusion holds, the score can stay low even when a crisis is building. A reported fever, breathing difficulty or chest pain lifts the cap. In case of severe pain, fever or shortness of breath, call 144 or the care team without waiting.

This score is not a validated medical device. It points toward hydration, rest and contact with the care team.